You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@jena.apache.org by Andy Seaborne <an...@apache.org> on 2012/06/19 16:30:12 UTC
Re: [] [] () TDB Transactional error logged by Fuseki server for
all update operations
Simon,
As someone who did not use 0.9.0, do you not think that we ought to move
cautiously?
Fixing the JENA-256 is a more major investigation and having fixed it
needs testing, if indeed, it is a regression rather than being more
careful, correctly, about the data - i.e. 0.9.0 may have been wrong.
The workaround for JENA-260 has been confirmed, and hopefully the proper
fix will be soon too.
Andy
On 19/06/12 15:23, Simon Helsen (JIRA) wrote:
>
> [ https://issues.apache.org/jira/browse/JENA-260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13396795#comment-13396795 ]
>
> Simon Helsen commented on JENA-260:
> -----------------------------------
>
> right, like the performance regression in JENA-256 (that will be quite visible for clients who came from 0.9.0)
Re: [] [] () TDB Transactional error logged by Fuseki server for
all update operations
Posted by Andy Seaborne <an...@apache.org>.
On 19/06/12 16:40, Simon Helsen wrote:
> Right now, it just startling that in the span of a few weeks, a
> serious performance regression got introduced and that it is not
> understood.
The choice of language is yours.
> On a side note: you seem to believe that 256 will be difficult to
> fix.
I said I think it needs a decent chunk of time to investigate; anyone
can do that, or start that or even poke around and see what's what.
It was reported last Thursday at 16:48 (BST) after several rounds of
email starting last Wednesday (15:06 BST) necessary to refine the issue
from the original report. Likely, any change will need more than
functional testing.
So, two, nearly three, whole "working" days so far.
Andy
Re: [] [] () TDB Transactional error logged by Fuseki server for all update
operations
Posted by Simon Helsen <sh...@ca.ibm.com>.
Andy,
I was not suggesting to wait. That is a misunderstanding. I understand
that releasing often has benefits. I just hope that principle will also
apply once 256 is fixed (assuming it can be fixed). I.e. a quick release
especially if the performance regression was unnecessary.
On a side note: you seem to believe that 256 will be difficult to fix.
That may be, but I am not sure if it _has_ to be difficult. The
performance difference between May 15 and the release is so big that there
must be something obvious that caused it (I am saying this intuitively, I
can't prove this obviously). If that thing cannot easily be changed then
so be it. At least, it is understood why the regression was introduced.
Right now, it just startling that in the span of a few weeks, a serious
performance regression got introduced and that it is not understood.
As for IBM teams using Jena, yes, I am sure it is used a lot. I don't know
all of them and some are more open than others. As you can imagine, IBM
has 400k employees and we are not exactly one organization. IBM teams will
not quickly go open with their adoption, because it may require
permissions or other loopholes. I can't do anything about that. I know
that in my group, we can communicate and even provide limited fixes
(always requires review and approval though). If some teams in IBM use TDB
as the performance baseline then you should see that as a compliment. It
is probably because Jena/TDB is the most significant open source
RDF/Sparql implementation right now.
Simon
From:
Andy Seaborne <an...@apache.org>
To:
Simon Helsen/Toronto/IBM@IBMCA
Cc:
dev@jena.apache.org
Date:
06/19/2012 11:10 AM
Subject:
Re: [] [] () TDB Transactional error logged by Fuseki server for all
update operations
On 19/06/12 15:50, Simon Helsen wrote:
> Andy,
>
> I was basing that conclusion on the observations by Laurent. While I
> understand that it is possible 0.9.0 had actual functional problems
> (allowing it to be "faster"), the logic I am employing goes like this:
Release twice?
> 1) Laurent finds that 0.9.0 is considerably faster than 0.9.1 release
Probably - last we heard, it was a read/write mix. Be good to
understand the result properly; it will save a large amount of time
investigating and also provide a second test case.
> 2) I find that 0.9.1 (May 15) is considerably faster than 0.9.1 release
> (reported in Jena-256)
> 3) we did significant performance and functional testing (with massive
> concurrency etc.) with the 0.9.1 May 15 build which showed by our
> standards that 0.9.1 May 15 appears functionally correct from our point
> of view
> 4) my past experience with previous tx problems shows me that our
> internal test coverage is quite extensive
> 5) there is another team here I loosely collaborate with who has been
> using and functionally testing 0.9.0 for a while now and they have not
> reported any problems
Yes - several groups in IBM use Jena; some say some publicly, some
don't. Seems like there are more TDB users than I was aware of. Maybe
they will de-cloak and report their experiences.
> Now, to point 4), of course, it is biased. We use Jena in a specific
> manner and thus we miss a large class of functional testing that does
> not concern us. So, yes, it is possible that whatever happened after May
> 15 was important to fix functional problems. And of course, it is also
> quite possible there were or even are other problems we are simply not
> yet aware of.
It's not the only issue that, given infinite resources, would be fixed.
> I guess I am trying to point out that performance could be important
> enough to produce a new release as well as it may keep people from
> upgrading if they find the new release is visibly slower without
> significant other gains (for them)
err - as do log files with nasty looking exceptions and bad handling of
transaction 0.2.1 release. That can be fixed now.
Fuseki 0.2.2 is better than 0.2.1; otherwise people are using snapshots
to work round data robustness issues.
Performance of TDB is something that other groups in IBM has been keen
to mention in public presentations.
I confess I don't really see your overall argument to force a wait
rather than have several releases, each of which reduces the support
costs each time.
Andy
>
> Simon
>
>
> From: Andy Seaborne <an...@apache.org>
> To: dev@jena.apache.org
> Date: 06/19/2012 10:31 AM
> Subject: Re: [] [] () TDB Transactional error logged by
Fuseki server
> for all update operations
>
>
> ------------------------------------------------------------------------
>
>
>
> Simon,
>
> As someone who did not use 0.9.0, do you not think that we ought to move
> cautiously?
>
> Fixing the JENA-256 is a more major investigation and having fixed it
> needs testing, if indeed, it is a regression rather than being more
> careful, correctly, about the data - i.e. 0.9.0 may have been wrong.
>
> The workaround for JENA-260 has been confirmed, and hopefully the proper
> fix will be soon too.
>
> Andy
>
> On 19/06/12 15:23, Simon Helsen (JIRA) wrote:
> >
> > [
>
https://issues.apache.org/jira/browse/JENA-260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13396795#comment-13396795
> <
https://issues.apache.org/jira/browse/JENA-260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13396795#comment-13396795
>]
> >
> > Simon Helsen commented on JENA-260:
> > -----------------------------------
> >
> > right, like the performance regression in JENA-256 (that will be
> quite visible for clients who came from 0.9.0)
>
>
>
Re: [] [] () TDB Transactional error logged by Fuseki server for
all update operations
Posted by Andy Seaborne <an...@apache.org>.
On 19/06/12 15:50, Simon Helsen wrote:
> Andy,
>
> I was basing that conclusion on the observations by Laurent. While I
> understand that it is possible 0.9.0 had actual functional problems
> (allowing it to be "faster"), the logic I am employing goes like this:
Release twice?
> 1) Laurent finds that 0.9.0 is considerably faster than 0.9.1 release
Probably - last we heard, it was a read/write mix. Be good to
understand the result properly; it will save a large amount of time
investigating and also provide a second test case.
> 2) I find that 0.9.1 (May 15) is considerably faster than 0.9.1 release
> (reported in Jena-256)
> 3) we did significant performance and functional testing (with massive
> concurrency etc.) with the 0.9.1 May 15 build which showed by our
> standards that 0.9.1 May 15 appears functionally correct from our point
> of view
> 4) my past experience with previous tx problems shows me that our
> internal test coverage is quite extensive
> 5) there is another team here I loosely collaborate with who has been
> using and functionally testing 0.9.0 for a while now and they have not
> reported any problems
Yes - several groups in IBM use Jena; some say some publicly, some
don't. Seems like there are more TDB users than I was aware of. Maybe
they will de-cloak and report their experiences.
> Now, to point 4), of course, it is biased. We use Jena in a specific
> manner and thus we miss a large class of functional testing that does
> not concern us. So, yes, it is possible that whatever happened after May
> 15 was important to fix functional problems. And of course, it is also
> quite possible there were or even are other problems we are simply not
> yet aware of.
It's not the only issue that, given infinite resources, would be fixed.
> I guess I am trying to point out that performance could be important
> enough to produce a new release as well as it may keep people from
> upgrading if they find the new release is visibly slower without
> significant other gains (for them)
err - as do log files with nasty looking exceptions and bad handling of
transaction 0.2.1 release. That can be fixed now.
Fuseki 0.2.2 is better than 0.2.1; otherwise people are using snapshots
to work round data robustness issues.
Performance of TDB is something that other groups in IBM has been keen
to mention in public presentations.
I confess I don't really see your overall argument to force a wait
rather than have several releases, each of which reduces the support
costs each time.
Andy
>
> Simon
>
>
> From: Andy Seaborne <an...@apache.org>
> To: dev@jena.apache.org
> Date: 06/19/2012 10:31 AM
> Subject: Re: [] [] () TDB Transactional error logged by Fuseki server
> for all update operations
>
>
> ------------------------------------------------------------------------
>
>
>
> Simon,
>
> As someone who did not use 0.9.0, do you not think that we ought to move
> cautiously?
>
> Fixing the JENA-256 is a more major investigation and having fixed it
> needs testing, if indeed, it is a regression rather than being more
> careful, correctly, about the data - i.e. 0.9.0 may have been wrong.
>
> The workaround for JENA-260 has been confirmed, and hopefully the proper
> fix will be soon too.
>
> Andy
>
> On 19/06/12 15:23, Simon Helsen (JIRA) wrote:
> >
> > [
> https://issues.apache.org/jira/browse/JENA-260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13396795#comment-13396795
> <https://issues.apache.org/jira/browse/JENA-260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13396795#comment-13396795>]
> >
> > Simon Helsen commented on JENA-260:
> > -----------------------------------
> >
> > right, like the performance regression in JENA-256 (that will be
> quite visible for clients who came from 0.9.0)
>
>
>
Re: [] [] () TDB Transactional error logged by Fuseki server for all update
operations
Posted by Simon Helsen <sh...@ca.ibm.com>.
Andy,
I was basing that conclusion on the observations by Laurent. While I
understand that it is possible 0.9.0 had actual functional problems
(allowing it to be "faster"), the logic I am employing goes like this:
1) Laurent finds that 0.9.0 is considerably faster than 0.9.1 release
2) I find that 0.9.1 (May 15) is considerably faster than 0.9.1 release
(reported in Jena-256)
3) we did significant performance and functional testing (with massive
concurrency etc.) with the 0.9.1 May 15 build which showed by our
standards that 0.9.1 May 15 appears functionally correct from our point of
view
4) my past experience with previous tx problems shows me that our internal
test coverage is quite extensive
5) there is another team here I loosely collaborate with who has been
using and functionally testing 0.9.0 for a while now and they have not
reported any problems
Now, to point 4), of course, it is biased. We use Jena in a specific
manner and thus we miss a large class of functional testing that does not
concern us. So, yes, it is possible that whatever happened after May 15
was important to fix functional problems. And of course, it is also quite
possible there were or even are other problems we are simply not yet aware
of.
I guess I am trying to point out that performance could be important
enough to produce a new release as well as it may keep people from
upgrading if they find the new release is visibly slower without
significant other gains (for them)
Simon
From:
Andy Seaborne <an...@apache.org>
To:
dev@jena.apache.org
Date:
06/19/2012 10:31 AM
Subject:
Re: [] [] () TDB Transactional error logged by Fuseki server for all
update operations
Simon,
As someone who did not use 0.9.0, do you not think that we ought to move
cautiously?
Fixing the JENA-256 is a more major investigation and having fixed it
needs testing, if indeed, it is a regression rather than being more
careful, correctly, about the data - i.e. 0.9.0 may have been wrong.
The workaround for JENA-260 has been confirmed, and hopefully the proper
fix will be soon too.
Andy
On 19/06/12 15:23, Simon Helsen (JIRA) wrote:
>
> [
https://issues.apache.org/jira/browse/JENA-260?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13396795#comment-13396795
]
>
> Simon Helsen commented on JENA-260:
> -----------------------------------
>
> right, like the performance regression in JENA-256 (that will be quite
visible for clients who came from 0.9.0)