I just tried Jason's script (modified it to use 500 000 docs instead
of 50 000) against 1.2.x and 1.1.1, using OTP R14B03. Here's my
results:
1.2.x:
$ port=5984 ./test.sh
"none"
Filling db.
done
HTTP/1.1 200 OK
Server: CouchDB/1.2.0 (Erlang OTP/R14B03)
Date: Mon, 27 Feb 2012 16:08:43 GMT
Content-Type: text/plain; charset=utf-8
Content-Length: 252
Cache-Control: must-revalidate
{"db_name":"db1","doc_count":500001,"doc_del_count":0,"update_seq":500001,"purge_seq":0,"compact_running":false,"disk_size":130494577,"data_size":130490673,"instance_start_time":"1330358830830086","disk_format_version":6,"committed_update_seq":500001}
Building view.
real 1m5.725s
user 0m0.006s
sys 0m0.005s
done
1.1.1:
$ port=5984 ./test.sh
""
Filling db.
done
HTTP/1.1 200 OK
Server: CouchDB/1.1.2a785d32f-git (Erlang OTP/R14B03)
Date: Mon, 27 Feb 2012 16:15:33 GMT
Content-Type: text/plain;charset=utf-8
Content-Length: 230
Cache-Control: must-revalidate
{"db_name":"db1","doc_count":500001,"doc_del_count":0,"update_seq":500001,"purge_seq":0,"compact_running":false,"disk_size":122142818,"instance_start_time":"1330359233327316","disk_format_version":5,"committed_update_seq":500001}
Building view.
real 1m4.249s
user 0m0.006s
sys 0m0.005s
done
I don't see any significant difference there.
Regarding COUCHDB-1186, the only thing that might cause some non
determinism and affect performance is the queing/dequeing. Depending
on timings, it's possible the writer is dequeing less items per
dequeue operation and therefore inserting smaller batches into the
btree. The following small change ensures larger batches (while still
respecting the queue max size/item count):
http://friendpaste.com/178nPFgfyyeGf2vtNRpL0w
Running the test with this change:
$ port=5984 ./test.sh
"none"
Filling db.
done
HTTP/1.1 200 OK
Server: CouchDB/1.2.0 (Erlang OTP/R14B03)
Date: Mon, 27 Feb 2012 16:23:20 GMT
Content-Type: text/plain; charset=utf-8
Content-Length: 252
Cache-Control: must-revalidate
{"db_name":"db1","doc_count":500001,"doc_del_count":0,"update_seq":500001,"purge_seq":0,"compact_running":false,"disk_size":130494577,"data_size":130490673,"instance_start_time":"1330359706846104","disk_format_version":6,"committed_update_seq":500001}
Building view.
real 0m49.762s
user 0m0.006s
sys 0m0.005s
done
If there's no objection, I'll push that patch.
Also, another note, I noticed sometime ago that with master, using OTP
R15B I got a performance drop of 10% to 15% compared to using master
with OTP R14B04. Maybe it applies to 1.2.x as well.
On Mon, Feb 27, 2012 at 5:33 AM, Robert Newson <rnew...@apache.org> wrote:
> Bob D, can you give more details on the data set you're testing?
> Number of docs, size/complexity of docs, etc? Basically, enough info
> that I could write a script to automate building an equivalent
> database.
>
> I wrote a quick bash script to make a database and time a view build
> here: http://friendpaste.com/7kBiKJn3uX1KiGJAFPv4nK
>
> B.
>
> On 27 February 2012 13:15, Jan Lehnardt <j...@apache.org> wrote:
>>
>> On Feb 27, 2012, at 12:58 , Bob Dionne wrote:
>>
>>> Thanks for the clarification. I hope I'm not conflating things by
>>> continuing the discussion here, I thought that's what you requested?
>>
>> The discussion we had on IRC was regarding collecting more data items for
>> the performance regression before we start to draw conclusions.
>>
>> My intention here is to understand what needs doing before we can release
>> 1.2.0.
>>
>> I'll reply inline for the other issues.
>>
>>> I just downloaded the release candidate again to start fresh. "make
>>> distcheck" hangs on this step:
>>>
>>> /Users/bitdiddle/Downloads/apache-couchdb-1.2.0/apache-couchdb-1.2.0/_build/../test/etap/150-invalid-view-seq.t
>>> ......... 6/?
>>>
>>> Just stops completely. This is on R15B which has been rebuilt to use the
>>> recommended older SSL version. I haven't looked into this crashing too
>>> closely but I'm suspicious that I only see it with couchdb and never with
>>> bigcouch and never using the 1.2.x branch from source or any branch for
>>> that matter
>>
>> From the release you should run `make check`, not make distcheck. But I
>> assume you see a hang there too, as I have and others (yet not everybody),
>> too. I can't comment on BigCouch and what is different there. It is
>> interesting that 1.2.x won't hang. For me, `make check` in 1.2.x on R15B
>> hangs sometimes, in different places. I'm currently trying to gather more
>> information about this.
>>
>> The question here is whether `make check` passing in R15B is a release
>> requirement. In my vote I considered no, but I am happy to go with a
>> community decision if it emerges. What is your take here?
>>
>> In addition, this just shouldn't be a question, so we should investigate why
>> this happens at all and address the issue, hence COUCHDB-1424. Any insight
>> here would be appreciated as well.
>>
>>
>>> In the command line tests, 2,7, 27, and 32 fail. but it differs from run to
>>> run.
>>
>> I assume you mean the JS tests. Again, this isn't supposed to work in 1.2.x.
>> I'm happy to backport my changes from master to 1.2.x to make that work, but
>> I refrained from that because I didn't want to bring too much change to a
>> release branch. I'm happy to reconsider, but I don't think a release vote is
>> a good place to discuss feature backports.
>>
>>
>>> On Chrome attachment_ranges fails and it hangs on replicator_db
>>
>> This one is an "explaining away", but I think it is warranted. Chrome is
>> broken for attachment_ranges. I don't know if we reported this upstream
>> (Robert N?), but this isn't a release blocker. For the replicator_db test,
>> can you try running that in other browsers. I understand it is not the best
>> of situation (hence the move to the cli test suite for master), but if you
>> get this test to pass in at least one other browsers, this isn't a problem
>> that holds 1.2.x.
>>
>>
>>> With respect to performance I think comparisons with 1.1.x are important. I
>>> think almost any use case, contrived or otherwise should not be dismissed
>>> as a pathological or edge case. Bob's script is as simple as it gets and to
>>> me is a great smoke test. We need to figure out the reason 1.2 is clearly
>>> slower in this case. If there are specific scenarios that 1.2.x is
>>> optimized for then we should document that and provide reasons for the
>>> trade-offs
>>
>> I want to make absolutely clear that I take any report of performance
>> regression very seriously. But I'm rather annoyed that no information about
>> this ends up on dev@. I understand that on IRC there's some shared
>> understanding of a few scenarios where performance regressions can be shown.
>> I asked three times now that these be posted to this mailing list. I'm not
>> asking for a comprehensive report, but anything really. I found Robert
>> Newson's simple test script on IRC and ran that to test a suspicion of mine
>> which I posted in an earlier mail (tiny docs -> slower, bigger docs ->
>> faster). Nobody else bothered to post this here. I see no discussion about
>> what is observed, what is expected, what would be acceptable for a release
>> of 1.2.0 as is and what not.
>>
>> As far as this list is concerned, we know that a few people claimed that
>> things are slower and it's very real and that we should hold the 1.2.0
>> release for it. I'm more than happy to hold the release until we figured out
>> the things I asked for above and help out figuring it all out. But we need
>> something to work with here.
>>
>> I also understand that this is a voluntary project and people don't have
>> infinite time to spend, but at least a message of "we're collecting things,
>> will report when done", would be *great* to start. So far we only have a
>> "hold the horses, there might be a something going on".
>>
>> Please let me know if this request is unreasonable or whether I am
>> overreacting.
>>
>> Sorry for the rant.
>>
>> To anyone who has been looking into performance regression, can you please
>> send to this list any info you have? If you have a comprehensive analysis,
>> awesome, if you just ran some script on a machine, just send us that, let's
>> collect all the data to get this situation solved! We need your help.
>>
>>
>> tl;dr:
>>
>> There's three issues at hand:
>>
>> - Robert D -1'd a release artefact. We want to understand what needs to
>> happen to make a release. This includes assessing the issues he raises and
>> squaring them against the release vote.
>>
>> - There's a vague (as far as dev@ is concerned) report about a performance
>> regression. We need to get behind that.
>>
>> - There's been a non-dev@ discussion about the performance regression and
>> that is referenced to influence a dev@ decision. We need that discussion's
>> information on dev@ to proceed.
>>
>>
>> And to make it absolutely clear again. The performance regression *is* an
>> issue and I am very grateful for the people, including Robert Newson, Robert
>> Dionne and Jason Smith, who look into it. It's just that we need to treat
>> this as an issue and get all this info onto dev@ or into JRIA.
>>
>>
>> Cheers
>> Jan
>> --
>>
>>
>>
>>>
>>> Cheers,
>>>
>>> Bob
>>>
>>>
>>> On Feb 26, 2012, at 4:07 PM, Jan Lehnardt wrote:
>>>
>>>> Bob,
>>>>
>>>> thanks for your reply
>>>>
>>>> I wasn't implying we should try to explain anything away. All of these are
>>>> valid concerns, I just wanted to get a better understanding on where the
>>>> bit flips from +0 to -1 and subsequently, how to address that boundary.
>>>
>>>
>>>
>>>
>>>> Ideally we can just fix all of the things you mention, but I think it is
>>>> important to understand them in detail, that's why I was going into them.
>>>> Ultimately, I want to understand what we need to do to ship 1.2.0.
>>>>
>>>> On Feb 26, 2012, at 21:22 , Bob Dionne wrote:
>>>>
>>>>> Jan,
>>>>>
>>>>> I'm -1 based on all of my evaluation. I've spent a few hours on this
>>>>> release now yesterday and today. It doesn't really pass what I would call
>>>>> the "smoke test". Almost everything I've run into has an explanation:
>>>>>
>>>>> 1. crashes out of the box - that's R15B, you need to recompile SSL and
>>>>> Erlang (we'll note on release notes)
>>>>
>>>> Have we spent any time on figuring out what the trouble here is?
>>>>
>>>>
>>>>> 2. etaps hang running make check. Known issue. Our etap code is out of
>>>>> date, recent versions of etap don't even run their own unit tests
>>>>
>>>> I have seen the etap hang as well, and I wasn't diligent enough to report
>>>> it in JIRA, I have done so now (COUCHDB-1424).
>>>>
>>>>
>>>>> 3. Futon tests fail. Some are known bugs (attachment ranges in Chrome) .
>>>>> Both Chrome and Safari also hang
>>>>
>>>> Do you have more details on where Chrome and Safari hang? Can you try
>>>> their private browsing features, double/triple check that caches are
>>>> empty? Can you get to a situation where you get all tests succeeding
>>>> across all browsers, even if individual ones fail on one or two others?
>>>>
>>>>
>>>>> 4. standalone JS tests fail. Again most of these run when run by
>>>>> themselves
>>>>
>>>> Which ones?
>>>>
>>>>
>>>>> 5. performance. I used real production data *because* Stefan on user
>>>>> reported performance degradation on his data set. Any numbers are
>>>>> meaningless for a single test. I also ran scripts that BobN and Jason
>>>>> Smith posted that show a difference between 1.1.x and 1.2.x
>>>>
>>>> You are conflating an IRC discussion we've had into this thread. The
>>>> performance regression reported is a good reason to look into other
>>>> scenarios where we can show slowdowns. But we need to understand what's
>>>> happening. Just from looking at dev@ all I see is some handwaving about
>>>> some reports some people have done (Not to discourage any work that has
>>>> been done on IRC and user@, but for the sake of a release vote thread,
>>>> this related information needs to be on this mailing list).
>>>>
>>>> As I said on IRC, I'm happy to get my hands dirty to understand the
>>>> regression at hand. But we need to know where we'd draw a line and say
>>>> this isn't acceptable for a 1.2.0.
>>>>
>>>>
>>>>> 6. Reviewed patch pointed to by Jason that may be the cause but it's hard
>>>>> to say without knowing the code analysis that went into the changes. You
>>>>> can see obvious local optimizations that make good sense but those are
>>>>> often the ones that get you, without knowing the call counts.
>>>>
>>>> That is a point that wasn't included in your previous mail. It's great
>>>> that there is progress, thanks for looking into this!
>>>>
>>>>
>>>>> Many of these issues can be explained away, but I think end users will be
>>>>> less forgiving. I think we already struggle with view performance. I'm
>>>>> interested to see how others evaluate this regression.
>>>>> I'll try this seatoncouch tool you mention later to see if I can
>>>>> construct some more definitive tests.
>>>>
>>>> Again, I'm not trying to explain anything away. I want to get a shared
>>>> understanding of the issues you raised and where we stand on solving them
>>>> squared against the ongoing 1.2.0 release.
>>>>
>>>> And again: Thanks for doing this thorough review and looking into
>>>> performance issue. I hope with your help we can understand all these
>>>> things a lot better very soon :)
>>>>
>>>> Cheers
>>>> Jan
>>>> --
>>>>
>>>>
>>>>>
>>>>> Best,
>>>>>
>>>>> Bob
>>>>> On Feb 26, 2012, at 2:29 PM, Jan Lehnardt wrote:
>>>>>
>>>>>>
>>>>>> On Feb 26, 2012, at 13:58 , Bob Dionne wrote:
>>>>>>
>>>>>>> -1
>>>>>>>
>>>>>>> R15B on OS X Lion
>>>>>>>
>>>>>>> I rebuilt OTP with an older SSL and that gets past all the crashes
>>>>>>> (thanks Filipe). I still see hangs when running make check, though any
>>>>>>> particular etap that hangs will run ok by itself. The Futon tests never
>>>>>>> run to completion in Chrome without hanging and the standalone JS tests
>>>>>>> also have fails.
>>>>>>
>>>>>> What part of this do you consider the -1? Can you try running the JS
>>>>>> tests in Firefox and or Safari? Can you get all tests pass at least once
>>>>>> across all browsers? The cli JS suite isn't supposed to work, so that
>>>>>> isn't a criterion. I've seen the hang in make check for R15B while
>>>>>> individual tests run as well, but I don't consider this blocking. While
>>>>>> I understand and support the notion that tests shouldn't fail, period,
>>>>>> we gotta work with what we have and master already has significant
>>>>>> improvements. What would you like to see changed to not -1 this release?
>>>>>>
>>>>>>> I tested the performance of view indexing, using a modest 200K doc db
>>>>>>> with a large complex view and there's a clear regression between 1.1.x
>>>>>>> and 1.2.x Others report similar results
>>>>>>
>>>>>> What is a large complex view? The complexity of the map/reduce functions
>>>>>> is rarely an indicator of performance, it's usually input doc size and
>>>>>> output/emit()/reduce data size. How big are the docs in your test and
>>>>>> how big is the returned data? I understand the changes for 1.2.x will
>>>>>> improve larger-data scenarios more significantly.
>>>>>>
>>>>>> Cheers
>>>>>> Jan
>>>>>> --
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> On Feb 23, 2012, at 5:25 PM, Bob Dionne wrote:
>>>>>>>
>>>>>>>> sorry Noah, I'm in debug mode today so I don't care to start mucking
>>>>>>>> with my stack, recompiling erlang, etc...
>>>>>>>>
>>>>>>>> I did try using that build repeatedly and it crashes all the time. I
>>>>>>>> find it very odd and I had seen those before as I said on my older
>>>>>>>> macbook.
>>>>>>>>
>>>>>>>> I do see the hangs Jan describes in the etaps, they have been there
>>>>>>>> right along, so I'm confident this just the SSL issue. Why it only
>>>>>>>> happens on the build is puzzling, any source build of any branch works
>>>>>>>> just peachy.
>>>>>>>>
>>>>>>>> So I'd say I'm +1 based on my use of the 1.2.x branch but I'd like to
>>>>>>>> hear from Stefan, who reported the severe performance regression. BobN
>>>>>>>> seems to think we can ignore that, it's something flaky in that
>>>>>>>> fellow's environment. I tend to agree but I'm conservative
>>>>>>>>
>>>>>>>> On Feb 23, 2012, at 1:23 PM, Noah Slater wrote:
>>>>>>>>
>>>>>>>>> Can someone convince me this bus error stuff and segfaults is not a
>>>>>>>>> blocking issue.
>>>>>>>>>
>>>>>>>>> Bob tells me that he's followed the steps above and he's still
>>>>>>>>> experiencing
>>>>>>>>> the issues.
>>>>>>>>>
>>>>>>>>> Bob, you did follow the steps to install your own SSL right?
>>>>>>>>>
>>>>>>>>> On Thu, Feb 23, 2012 at 5:09 PM, Jan Lehnardt <j...@apache.org> wrote:
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On Feb 23, 2012, at 00:28 , Noah Slater wrote:
>>>>>>>>>>
>>>>>>>>>>> Hello,
>>>>>>>>>>>
>>>>>>>>>>> I would like call a vote for the Apache CouchDB 1.2.0 release,
>>>>>>>>>>> second
>>>>>>>>>> round.
>>>>>>>>>>>
>>>>>>>>>>> We encourage the whole community to download and test these
>>>>>>>>>>> release artifacts so that any critical issues can be resolved
>>>>>>>>>>> before the
>>>>>>>>>>> release is made. Everyone is free to vote on this release, so get
>>>>>>>>>>> stuck
>>>>>>>>>> in!
>>>>>>>>>>>
>>>>>>>>>>> We are voting on the following release artifacts:
>>>>>>>>>>>
>>>>>>>>>>> http://people.apache.org/~nslater/dist/1.2.0/
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> These artifacts have been built from the following tree-ish in Git:
>>>>>>>>>>>
>>>>>>>>>>> 4cd60f3d1683a3445c3248f48ae064fb573db2a1
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Please follow the test procedure before voting:
>>>>>>>>>>>
>>>>>>>>>>> http://wiki.apache.org/couchdb/Test_procedure
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Thank you.
>>>>>>>>>>>
>>>>>>>>>>> Happy voting,
>>>>>>>>>>
>>>>>>>>>> Signature and hashes check out.
>>>>>>>>>>
>>>>>>>>>> Mac OS X 10.7.3, 64bit, SpiderMonkey 1.8.0, Erlang R14B04: make check
>>>>>>>>>> works fine, browser tests in Safari work fine.
>>>>>>>>>>
>>>>>>>>>> Mac OS X 10.7.3, 64bit, SpiderMonkey 1.8.5, Erlang R14B04: make check
>>>>>>>>>> works fine, browser tests in Safari work fine.
>>>>>>>>>>
>>>>>>>>>> FreeBSD 9.0, 64bit, SpiderMonkey 1.7.0, Erlang R14B04: make check
>>>>>>>>>> works
>>>>>>>>>> fine, browser tests in Safari work fine.
>>>>>>>>>>
>>>>>>>>>> CentOS 6.2, 64bit, SpiderMonkey 1.8.5, Erlang R14B04: make check
>>>>>>>>>> works
>>>>>>>>>> fine, browser tests in Firefox work fine.
>>>>>>>>>>
>>>>>>>>>> Ubuntu 11.4, 64bit, SpiderMonkey 1.8.5, Erlang R14B02: make check
>>>>>>>>>> works
>>>>>>>>>> fine, browser tests in Firefox work fine.
>>>>>>>>>>
>>>>>>>>>> Ubuntu 10.4, 32bit, SpiderMonkey 1.8.0, Erlang R13B03: make check
>>>>>>>>>> fails in
>>>>>>>>>> - 076-file-compression.t: https://gist.github.com/1893373
>>>>>>>>>> - 220-compaction-daemon.t: https://gist.github.com/1893387
>>>>>>>>>> This on runs in a VM and is 32bit, so I don't know if there's
>>>>>>>>>> anything in
>>>>>>>>>> the tests that rely on 64bittyness or the R14B03. Filipe, I think you
>>>>>>>>>> worked on both features, do you have an idea?
>>>>>>>>>>
>>>>>>>>>> I tried running it all through Erlang R15B on Mac OS X 1.7.3, but a
>>>>>>>>>> good
>>>>>>>>>> way into `make check` the tests would just stop and hang. The last
>>>>>>>>>> time,
>>>>>>>>>> repeatedly in 160-vhosts.t, but when run alone, that test finished
>>>>>>>>>> in under
>>>>>>>>>> five seconds. I'm not sure what the issue is here.
>>>>>>>>>>
>>>>>>>>>> Despite the things above, I'm happy to give this a +1 if we put a
>>>>>>>>>> warning
>>>>>>>>>> about R15B on the download page.
>>>>>>>>>>
>>>>>>>>>> Great work all!
>>>>>>>>>>
>>>>>>>>>> Cheers
>>>>>>>>>> Jan
>>>>>>>>>> --
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
--
Filipe David Manana,
"Reasonable men adapt themselves to the world.
Unreasonable men adapt the world to themselves.
That's why all progress depends on unreasonable men."
