[jira] Created: (LUCENE-1693) AttributeSource/TokenStream API improvements

AttributeSource/TokenStream API improvements
--------------------------------------------

                 Key: LUCENE-1693
                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
             Project: Lucene - Java
          Issue Type: Improvement
          Components: Analysis
            Reporter: Michael Busch
            Assignee: Michael Busch
            Priority: Minor
             Fix For: 2.9


This patch makes the following improvements to AttributeSource and
TokenStream/Filter:

- removes the set/getUseNewAPI() methods (including the standard
  ones). Instead by default incrementToken() throws a subclass of
  UnsupportedOperationException. The indexer tries to call
  incrementToken() initially once to see if the exception is thrown;
  if so, it falls back to the old API.

- introduces interfaces for all Attributes. The corresponding
  implementations have the postfix 'Impl', e.g. TermAttribute and
  TermAttributeImpl. AttributeSource now has a factory for creating
  the Attribute instances; the default implementation looks for
  implementing classes with the postfix 'Impl'. Token now implements
  all 6 TokenAttribute interfaces.

- new method added to AttributeSource:
  addAttributeImpl(AttributeImpl). Using reflection it walks up in the
  class hierarchy of the passed in object and finds all interfaces
  that the class or superclasses implement and that extend the
  Attribute interface. It then adds the interface->instance mappings
  to the attribute map for each of the found interfaces.

- AttributeImpl now has a default implementation of toString that uses
  reflection to print out the values of the attributes in a default
  formatting. This makes it a bit easier to implement AttributeImpl,
  because toString() was declared abstract before.

- Cloning is now done much more efficiently in
  captureState. The method figures out which unique AttributeImpl
  instances are contained as values in the attributes map, because
  those are the ones that need to be cloned. It creates a single
  linked list that supports deep cloning (in the inner class
  AttributeSource.State). AttributeSource keeps track of when this
  state changes, i.e. whenever new attributes are added to the
  AttributeSource. Only in that case will captureState recompute the
  state, otherwise it will simply clone the precomputed state and
  return the clone. restoreState(AttributeSource.State) walks the
  linked list and uses the copyTo() method of AttributeImpl to copy
  all values over into the attribute that the source stream
  (e.g. SinkTokenizer) uses.

The cloning performance can be greatly improved if not multiple
AttributeImpl instances are used in one TokenStream. A user can
e.g. simply add a Token instance to the stream instead of the individual
attributes. Or the user could implement a subclass of AttributeImpl that
implements exactly the Attribute interfaces needed. I think this
should be considered an expert API (addAttributeImpl), as this manual
optimization is only needed if cloning performance is crucial. I ran
some quick performance tests using Tee/Sink tokenizers (which do
cloning) and the performance was roughly 20% faster with the new
API. I'll run some more performance tests and post more numbers then.

Note also that when we add serialization to the Attributes, e.g. for
supporting storing serialized TokenStreams in the index, then the
serialization should benefit even significantly more from the new API
than cloning.

Also, the TokenStream API does not change, except for the removal
of the set/getUseNewAPI methods. So the patches in LUCENE-1460
should still work.

All core tests pass, however, I need to update all the documentation
and also add some unit tests for the new AttributeSource
functionality. So this patch is not ready to commit yet, but I wanted
to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

     [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Michael Busch updated LUCENE-1693:
----------------------------------

    Attachment: lucene-1693.patch

Patch that includes all mentioned improvements, but needs cleanup, documentation improvements and junits.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720031#action_12720031 ]

Uwe Schindler commented on LUCENE-1693:
---------------------------------------

Why do you add a new class "SmallToken"? I think it should be the good old deprecated "Token".

In my opinion, I would not deprecate the old Token, instead the default factory should always create Token instead of all these default *Impl attributes.

What was your concusion about my idea yesterday, to pass the Token around even with the old API and copy it on demand, if return value by next() is not the same? In this case, Token should be the only implementation of all standard attributes.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720046#action_12720046 ]

Michael Busch commented on LUCENE-1693:
---------------------------------------

{quote}
Why do you add a new class "SmallToken"? I think it should be the good old deprecated "Token".
{quote}

I should have added a "no commit" comment to SmallToken. I just used it for performance tests, we don't have to commit it. It's an example of what an advanced user could do to speed up cloning. Token is still there and unchanged.

{quote}
In my opinion, I would not deprecate the old Token, instead the default factory should always create Token instead of all these default *Impl attributes.
{quote}

Yeah we shouldn't deprecate Token. Not sure about the default. It kind of depends on the use case. Cloning and memory is faster I think with only two or three Attributes compared to using the Token cloning is faster. But it'd be simpler to not have all the Impl classes... hmm not sure...

{quote}
What was your concusion about my idea yesterday, to pass the Token around even with the old API and copy it on demand
{quote}

I don't think the indexer should know at all about Token? It should only use the interfaces so that we can maintain the full flexibility.
Also I don't really like the fact very much that some user might get a performance hit. I had the idea to throw the exception in incrementToken() to automatically being able to fallback to the old API. I think this is nice and gets rid of the explicit useNewAPI methods. The only drawback is still the fact that we have to implement both old and new APIs in Lucene's tokenizers and filters until we remove the old API. Grant won't like this :) but I think this is better than the possible performance hit? Also we don't add new filters THAT often to Lucene and implementing both methods is often mostly copy&paste.

Do you like the idea with the UnsupportedOperationException?

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720052#action_12720052 ]

Uwe Schindler commented on LUCENE-1693:
---------------------------------------

{quote}
bq. What was your concusion about my idea yesterday, to pass the Token around even with the old API and copy it on demand

I don't think the indexer should know at all about Token? It should only use the interfaces so that we can maintain the full flexibility.
Also I don't really like the fact very much that some user might get a performance hit. I had the idea to throw the exception in incrementToken() to automatically being able to fallback to the old API. I think this is nice and gets rid of the explicit useNewAPI methods. The only drawback is still the fact that we have to implement both old and new APIs in Lucene's tokenizers and filters until we remove the old API. Grant won't like this  but I think this is better than the possible performance hit? Also we don't add new filters THAT often to Lucene and implementing both methods is often mostly copy&paste.
{quote}

I did not mean to use Token in the indexer, I would like to remove the whole old API from Token (even the UOE). My idea would be the following:
- Let the default interfaces implemented by Token, the default factory creates it for all requests to the default attributes
- If somebody implements the new API, the indexer can use it without problems. If he doesn't, the default impl in TokenStream would call next(Token), using the token instance from AttributeSource. If the method returns with another Token instance (because it did not reuse, which is seldom I think), copy this returned token into the per instance AttributeSource Token.
- The other way round, if one uses a TokenStream with the old API (own indexer, query parser,...), the TokenStream only implemented the new API, the deprectated old method would also have a default impl, ignoring the token supplied to next() and returning always the instance-token after calling incrementToken().

Because of this, the indexer would always use the new API, the old API is wrapped. Core analyzers only need to implement the new methods (or could even stay with the old).

There are two problems:
- If somebody does not implement either method, the call to incrementToken() will loop endless, there should be some early break with UOE in this case. Do not know how to implement this without inspecting the stack trace in the default methods.
- Filters are a bit tricky: They could pass in both methods per default to the delegate. The problem, if one only implements one of the methods, the other passes down to the delegate, and doing nothing. But this could be fixed by delegating in the deprecated old method always to the new one (or vice versa).

Hope this was clear, maybe I should create a patch.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720052#action_12720052 ]

Uwe Schindler edited comment on LUCENE-1693 at 6/16/09 3:43 AM:
----------------------------------------------------------------

{quote}
bq. What was your concusion about my idea yesterday, to pass the Token around even with the old API and copy it on demand

I don't think the indexer should know at all about Token? It should only use the interfaces so that we can maintain the full flexibility.
Also I don't really like the fact very much that some user might get a performance hit. I had the idea to throw the exception in incrementToken() to automatically being able to fallback to the old API. I think this is nice and gets rid of the explicit useNewAPI methods. The only drawback is still the fact that we have to implement both old and new APIs in Lucene's tokenizers and filters until we remove the old API. Grant won't like this  but I think this is better than the possible performance hit? Also we don't add new filters THAT often to Lucene and implementing both methods is often mostly copy&paste.
{quote}

I did not mean to use Token in the indexer, I would like to remove the whole old API from the indexer (even the UOE). My idea would be the following:
- Let the default interfaces implemented by Token, the default factory creates it for all requests to the default attributes
- If somebody implements the new API, the indexer can use it without problems. If he doesn't, the default impl in TokenStream would call next(Token), using the token instance from AttributeSource. If the method returns with another Token instance (because it did not reuse, which is seldom I think), copy this returned token into the per instance AttributeSource Token.
- The other way round, if one uses a TokenStream with the old API (own indexer, query parser,...), the TokenStream only implemented the new API, the deprectated old method would also have a default impl, ignoring the token supplied to next() and returning always the instance-token after calling incrementToken().

Because of this, the indexer would always use the new API, the old API is wrapped. Core analyzers only need to implement the new methods (or could even stay with the old).

There are two problems:
- If somebody does not implement either method, the call to incrementToken() will loop endless, there should be some early break with UOE in this case. Do not know how to implement this without inspecting the stack trace in the default methods.
- Filters are a bit tricky: They could pass in both methods per default to the delegate. The problem, if one only implements one of the methods, the other passes down to the delegate, and doing nothing. But this could be fixed by delegating in the deprecated old method always to the new one (or vice versa).

Hope this was clear, maybe I should create a patch.

      was (Author: thetaphi):
    {quote}
bq. What was your concusion about my idea yesterday, to pass the Token around even with the old API and copy it on demand

I don't think the indexer should know at all about Token? It should only use the interfaces so that we can maintain the full flexibility.
Also I don't really like the fact very much that some user might get a performance hit. I had the idea to throw the exception in incrementToken() to automatically being able to fallback to the old API. I think this is nice and gets rid of the explicit useNewAPI methods. The only drawback is still the fact that we have to implement both old and new APIs in Lucene's tokenizers and filters until we remove the old API. Grant won't like this  but I think this is better than the possible performance hit? Also we don't add new filters THAT often to Lucene and implementing both methods is often mostly copy&paste.
{quote}

I did not mean to use Token in the indexer, I would like to remove the whole old API from Token (even the UOE). My idea would be the following:
- Let the default interfaces implemented by Token, the default factory creates it for all requests to the default attributes
- If somebody implements the new API, the indexer can use it without problems. If he doesn't, the default impl in TokenStream would call next(Token), using the token instance from AttributeSource. If the method returns with another Token instance (because it did not reuse, which is seldom I think), copy this returned token into the per instance AttributeSource Token.
- The other way round, if one uses a TokenStream with the old API (own indexer, query parser,...), the TokenStream only implemented the new API, the deprectated old method would also have a default impl, ignoring the token supplied to next() and returning always the instance-token after calling incrementToken().

Because of this, the indexer would always use the new API, the old API is wrapped. Core analyzers only need to implement the new methods (or could even stay with the old).

There are two problems:
- If somebody does not implement either method, the call to incrementToken() will loop endless, there should be some early break with UOE in this case. Do not know how to implement this without inspecting the stack trace in the default methods.
- Filters are a bit tricky: They could pass in both methods per default to the delegate. The problem, if one only implements one of the methods, the other passes down to the delegate, and doing nothing. But this could be fixed by delegating in the deprecated old method always to the new one (or vice versa).

Hope this was clear, maybe I should create a patch.
 

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720060#action_12720060 ]

Michael Busch commented on LUCENE-1693:
---------------------------------------

{quote}
- If somebody implements the new API, the indexer can use it without problems. If he doesn't, the default impl in TokenStream would call next(Token), using the token instance from AttributeSource. If the method returns with another Token instance (because it did not reuse, which is seldom I think), copy this returned token into the per instance AttributeSource Token.
{quote}

What if you currently have a filter that caches the token instance passed into next(Token) and returns a different one. If you then copy the values from the returned token into the per instance Attribute Token, you alter the cached one, because it's the same instance.

Not sure if this is common or even allowed, but these are the sideeffects I'm worried about.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720066#action_12720066 ]

Uwe Schindler commented on LUCENE-1693:
---------------------------------------

{quote}
What if you currently have a filter that caches the token instance passed into next(Token) and returns a different one. If you then copy the values from the returned token into the per instance Attribute Token, you alter the cached one, because it's the same instance.

Not sure if this is common or even allowed, but these are the sideeffects I'm worried about.
{quote}

See:
http://lucene.apache.org/java/2_4_1/api/org/apache/lucene/analysis/TokenStream.html

It states in next(Token): Also, the producer must make no assumptions about a Token after it has been returned: the caller may arbitrarily change it. If the producer needs to hold onto the token for subsequent calls, it must clone() it before storing it. Note that a TokenFilter is considered a consumer.

So I see no problem with this.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720076#action_12720076 ]

Shai Erera commented on LUCENE-1693:
------------------------------------

I have a couple of TokenFilters that work that way - get a Token from the wrapped TokenStream, cache it (cloning) and return another Token instead. Especially as the Tokenizer fills the Token w/ characters, I think that cloning is the only option if you want to hold on to a Token that was returned by a wrapped TokenStream.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720081#action_12720081 ]

Uwe Schindler commented on LUCENE-1693:
---------------------------------------

If you clone, you would not fall into the mentioned problem. The reuseable Token passed to next() is owned by the caller.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720196#action_12720196 ]

Michael Busch commented on LUCENE-1693:
---------------------------------------

But, the additional copying would affect performance in Shai's case. The same with CachingTokenFilter and Tee/Sink filer/tokenizer. I don't think anymore it's such a corner case to not return the reusable token.

From a user's perspective, the UnsupportedOperationException solution and your solution should not be distinguishable. The advantage of your solution is cleaner, less cluttered code, and it removes the burden from the developers to implement both APIs in the transition period.The disadvantage is a possible performance hit in the cases mentioned above.
The advantage of the UOE solution is no performance hit for users of the old API, the disadvantage the need to implement both APIs.

Even though I think your solution is very elegant, Uwe, I'm in favor of the one that doesn't affect performance significantly. What do you or others think about the possible performance hit? Not a big deal? Or worth keeping the old code around for a while?

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720222#action_12720222 ]

Michael McCandless commented on LUCENE-1693:
--------------------------------------------

bq. What do you or others think about the possible performance hit?

I haven't been able to keep up with this, but this one caught my eye!

I think performance of analysis, especially the common case (all my tokenizers are new, I'm not cloning, etc.) is important.  It sounds like this performance hit was specifically on cloning... but still we should try not to lose performance if we can help it.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720227#action_12720227 ]

Grant Ingersoll commented on LUCENE-1693:
-----------------------------------------

I think cloning is more prevalent than people think.  For instance, the RemoveDuplicatesTokenFilter in Solr is in the default schema and is thus used, albeit naively, by a whole lot of people.  Furthermore, I've seen quite a few apps that need to buffer tokens before spitting them out, things like phrase detection, n-grams, entity extraction, sentence detection, part of speech detection, parsing, etc.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720230#action_12720230 ]

Shai Erera commented on LUCENE-1693:
------------------------------------

Perhaps I'm missing something, but I don't understand what is the performance hit that's been discussed. If you don't need to clone a Token in order to cache it, don't do it and I agree we shouldn't require it or anything.

In my scenario, I do need to cache a Token, and then on subsequent calls I use it to generate new Tokens. In next(Token) I populate the given Token with information from the cached Token. That goes until it's consumed, and then I call super.next(Token), and copy its content into the cached Token (I don't use clone() since that allocates a new Token which i don't need).

I don't suffer in my scenario, I'm quite happy w/ how it works and expect to be able to do the exact same thing when I switch to the new API. If that will be the case, then it's fine with me.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720236#action_12720236 ]

Michael Busch commented on LUCENE-1693:
---------------------------------------

{quote}
Perhaps I'm missing something, but I don't understand what is the performance hit that's been discussed.
{quote}

We're discussing here basically how we want to treat backwards-compatibility for the old API. In the patch attached here the default implementation of TokenStream.incrementToken() throws a UnsupportedOperationException. If the consumer, i.e. indexer, hits that exception it knows to fallback to the old API. This change has the same disadvantage as trunk currently has: for all streams/filters in the core and contrib we have to implement both the old next(Token) and the new incrementToken().

To avoid this disadvantage Uwe is proposing an elegant change: each TokenStreams owns a reusableToken, which is shared with the consumer (indexer). That's easily possible, because the patch here changes Token to implement all TokenAttribute interfaces. The default implementation of incrementToken() then calls next(Token) and compares the Token next(Token) returns with the reusableToken. If it's not identical then we need to copy all values from the returned Token into the reusableToken, so that the consumer gets the values. This additional copying is what concerns me in terms of performance.
Currently e.g. SinkTokenizer always returns a different Token, which means we would always have to copy. We could change SinkTokenizer to not clone, but call Token.reinit(), which copies. But I think there might be a significant amount of filters out there that do similar things.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720237#action_12720237 ]

Uwe Schindler commented on LUCENE-1693:
---------------------------------------

Grant: But with the new TokenStream API, you must always clone the attributes before caching, because the TokenFilter only has one "working" instance that changes all the time.
In this case it would not be better than wrapping next(Token) with the new api and feeding in the current instance token and copy it back into the instance-one if next(Token) returns a different instance.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720237#action_12720237 ]

Uwe Schindler edited comment on LUCENE-1693 at 6/16/09 9:57 AM:
----------------------------------------------------------------

Grant: But with the new TokenStream API, you must always clone the attributes before caching, because the TokenFilter only has *one* final "working" instance per Attribute whose contents change all the time.
In this case it would not be better than wrapping next(Token) with the new api and feeding in the current instance token and copy it back into the instance-one if next(Token) returns a different instance.

      was (Author: thetaphi):
    Grant: But with the new TokenStream API, you must always clone the attributes before caching, because the TokenFilter only has one "working" instance that changes all the time.
In this case it would not be better than wrapping next(Token) with the new api and feeding in the current instance token and copy it back into the instance-one if next(Token) returns a different instance.
 

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720239#action_12720239 ]

Uwe Schindler commented on LUCENE-1693:
---------------------------------------

bq. Currently e.g. SinkTokenizer always returns a different Token, which means we would always have to copy. We could change SinkTokenizer to not clone, but call Token.reinit(), which copies. But I think there might be a significant amount of filters out there that do similar things.

If SinkTokenizer does this, it would not be very fast even with the old API because it creates a lot of new instances instead of reusing? So the problem here is in SinkTokenizer.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720240#action_12720240 ]

Shai Erera commented on LUCENE-1693:
------------------------------------

Thanks Michael. Given that, I don't think my private scenario will be affected in terms of performance, since my cached token is a different instance then the one that's been passed on the filters, and I just populate the instance passed on the filters with data from the cached token. So besides me copying char[], the consumer will still share the same instance.

Perhaps we should document that as a best practice or something, in light of this proposal?

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...

    [ https://issues.apache.org/jira/browse/LUCENE-1693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12720244#action_12720244 ]

Grant Ingersoll commented on LUCENE-1693:
-----------------------------------------

Right, SinkTokenizer is only really cost effective when the number of Tokens that are "teed" is roughly less than 1/2, according to my basic tests.  After that, just doing a copy field approach is fine.

> AttributeSource/TokenStream API improvements
> --------------------------------------------
>
>                 Key: LUCENE-1693
>                 URL: https://issues.apache.org/jira/browse/LUCENE-1693
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Analysis
>            Reporter: Michael Busch
>            Assignee: Michael Busch
>            Priority: Minor
>             Fix For: 2.9
>
>         Attachments: lucene-1693.patch
>
>
> This patch makes the following improvements to AttributeSource and
> TokenStream/Filter:
> - removes the set/getUseNewAPI() methods (including the standard
>   ones). Instead by default incrementToken() throws a subclass of
>   UnsupportedOperationException. The indexer tries to call
>   incrementToken() initially once to see if the exception is thrown;
>   if so, it falls back to the old API.
> - introduces interfaces for all Attributes. The corresponding
>   implementations have the postfix 'Impl', e.g. TermAttribute and
>   TermAttributeImpl. AttributeSource now has a factory for creating
>   the Attribute instances; the default implementation looks for
>   implementing classes with the postfix 'Impl'. Token now implements
>   all 6 TokenAttribute interfaces.
> - new method added to AttributeSource:
>   addAttributeImpl(AttributeImpl). Using reflection it walks up in the
>   class hierarchy of the passed in object and finds all interfaces
>   that the class or superclasses implement and that extend the
>   Attribute interface. It then adds the interface->instance mappings
>   to the attribute map for each of the found interfaces.
> - AttributeImpl now has a default implementation of toString that uses
>   reflection to print out the values of the attributes in a default
>   formatting. This makes it a bit easier to implement AttributeImpl,
>   because toString() was declared abstract before.
> - Cloning is now done much more efficiently in
>   captureState. The method figures out which unique AttributeImpl
>   instances are contained as values in the attributes map, because
>   those are the ones that need to be cloned. It creates a single
>   linked list that supports deep cloning (in the inner class
>   AttributeSource.State). AttributeSource keeps track of when this
>   state changes, i.e. whenever new attributes are added to the
>   AttributeSource. Only in that case will captureState recompute the
>   state, otherwise it will simply clone the precomputed state and
>   return the clone. restoreState(AttributeSource.State) walks the
>   linked list and uses the copyTo() method of AttributeImpl to copy
>   all values over into the attribute that the source stream
>   (e.g. SinkTokenizer) uses.
> The cloning performance can be greatly improved if not multiple
> AttributeImpl instances are used in one TokenStream. A user can
> e.g. simply add a Token instance to the stream instead of the individual
> attributes. Or the user could implement a subclass of AttributeImpl that
> implements exactly the Attribute interfaces needed. I think this
> should be considered an expert API (addAttributeImpl), as this manual
> optimization is only needed if cloning performance is crucial. I ran
> some quick performance tests using Tee/Sink tokenizers (which do
> cloning) and the performance was roughly 20% faster with the new
> API. I'll run some more performance tests and post more numbers then.
> Note also that when we add serialization to the Attributes, e.g. for
> supporting storing serialized TokenStreams in the index, then the
> serialization should benefit even significantly more from the new API
> than cloning.
> Also, the TokenStream API does not change, except for the removal
> of the set/getUseNewAPI methods. So the patches in LUCENE-1460
> should still work.
> All core tests pass, however, I need to update all the documentation
> and also add some unit tests for the new AttributeSource
> functionality. So this patch is not ready to commit yet, but I wanted
> to post it already for some feedback.

--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@...
For additional commands, e-mail: java-dev-help@...