incubator-couchdb-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stephen Bartell <snbart...@gmail.com>
Subject Re: replicating docs with tons of conflicts
Date Fri, 15 Mar 2013 17:40:46 GMT

On Mar 14, 2013, at 3:36 PM, Robert Newson <rnewson@apache.org> wrote:

> Runaway processes are the very devil but the problem is not specific
> to CouchDB, there is no CouchDB mechanism for this just as there's no
> bash/python/ruby/perl method to limit a while(true){} loop.

Totally makes sense.

> 
> Highly conflicted documents are painful to update and read. I can't do
> anything about that today.

Thanks for your feedback!

> 
> B.
> 
> On 14 March 2013 17:23, Stephen Bartell <snbartell@gmail.com> wrote:
>> Robert, this only works if I don't need to keep those docs around anymore.  In my
case, I want to keep the docs.  I don't want to keep the conflicts of the docs. Most importantly
thought, even if I delete all the conflicts on all my docs, I still have the problem of _deleted_docs.
 What I've seen is that only a few docs with a few thousand _deleted_docs each will plug up
Couch and render unusable. You can't get rid of it through natural means.
>> 
>> This is what Riyad was bringing up and what Ive implemented.  I have a program which
replicates from the troubled database _changes with the query param style=main_only.  This
allows me to still have the revision tree of the troubled database, but without the _deleted_conflicts.
 I can then wipe out the troubled db, recreate it, and replicate the shiny clean data back
into it.
>> 
>> This is unnatural and requires custom code to make happen.  I can live with it until
a better solution comes around.
>> 
>> What I'm really concerned about is how couchdb eats all my cpu.
>> 
>> Is there any way to ration the resources that couchdb uses? Like tell it not to use
more than 50% or something.  I think that couch eating all the resources on a machine just
because its reading loads of data is a bug.  Is this a reasonable conclusion?
>> 
>> On Mar 14, 2013, at 2:18 PM, Robert Newson <rnewson@apache.org> wrote:
>> 
>>> One trick, you can delete the doc and replicate with a filter like
>>> 'return !doc['_deleted'];' that blocks all deletes. the target db will
>>> then not receive any trace of these highly conflicted docs.
>>> 
>>> On 14 March 2013 14:10, Stephen Bartell <snbartell@gmail.com> wrote:
>>>> 
>>>> On Mar 14, 2013, at 11:44 AM, Robert Newson <rnewson@apache.org> wrote:
>>>> 
>>>>> Conflicts are *not* removed during compaction, CouchDB has no way of
>>>>> knowing which ones it would be ok to delete.
>>>> 
>>>> Yep, they need to be deleted in the context of the person/process manipulating
the docs.
>>>> 
>>>>> 
>>>>> CouchDB does struggle to process documents with lots of conflicts,
>>>>> we've encountered this at Cloudant a fair bunch. We resolve the
>>>>> conflicts via http if possible or, if that consistently fails, with a
>>>>> direct erlang manipulation. It's certainly something we need to
>>>>> improve.
>>>>> 
>>>> 
>>>> But even deleting them yields the same problem.  When replicating, the _deleted_conflicts
is carried over.
>>>> Users could be diligent in deleting conflicts, but still end up unable to
replicate their docs because of the volume of _deleted_conflicts.
>>>> 
>>>> Robert, thanks for chiming in.  I feel better knowing I'm in good company
with this problem. When this mine eventually goes off, couchdb is rendered useless because
beam.smp takes all the cpu.  Is there any way to ration the resources couchdb consumes?
>>>> 
>>>>> B.
>>>>> 
>>>>> On 14 March 2013 13:09, Riyad Kalla <rkalla@gmail.com> wrote:
>>>>>> Stephen,
>>>>>> I am probably wrong here (someone hop in and correct me), but I thought
>>>>>> Compaction would remove the old revisions (and conflicts) of docs.
>>>>>> 
>>>>>> Alternatively a question for the Couch devs, if Stephen set _revs_limit
to
>>>>>> something artifically low, say 1, and restarted couch and did a compaction,
>>>>>> would that force the DB to smash down the datastore to 1 rev per
doc and
>>>>>> remove the long-tail off these docs?
>>>>>> 
>>>>>> REF: http://wiki.apache.org/couchdb/Compaction
>>>>>> 
>>>>>> On Thu, Mar 14, 2013 at 2:02 AM, Stephen Bartell <snbartell@gmail.com>wrote:
>>>>>> 
>>>>>>> Hi all,
>>>>>>> 
>>>>>>> tldr; I've got a database with just a couple docs.  Conflict
management
>>>>>>> went unchecked and these docs have thousands of conflicts each.
>>>>>>> Replication fails.  Couch consumes all the server's cpu.
>>>>>>> 
>>>>>>> First the story, then the questions.  Please bear with me!
>>>>>>> 
>>>>>>> I wanted to replicate this database to another, new database.
 So I
>>>>>>> started the replication.  beam.smp took 100% of my cpu and the
replicator
>>>>>>> status held steady at a constant percent for quite a while. 
It eventually
>>>>>>> finished.
>>>>>>> 
>>>>>>> I thought maybe I should handle the conflicts and then replicate.
>>>>>>> Hopefuly it'll go faster next time.  So I cleared all the conflicts.
 I
>>>>>>> replicated again but this time I could not get anything to replicate.
>>>>>>> Again, cpu held steady, topped out. I eventually restarted couch.
>>>>>>> 
>>>>>>> I dug throughout the logs and saw that the POSTS were failing.
 I figure
>>>>>>> that the replicator was timing out when trying to post to couch.
>>>>>>> 
>>>>>>> I have a replicator that I've been working on thats written in
node.js.
>>>>>>> So I started that one up to do the same thing.  I drew inspiration
from
>>>>>>> Pouchdb's replicator and from Jens Alkes amazing replication
algorithm
>>>>>>> documentation, so my replicator follows more or less the same
story.  1)
>>>>>>> consume _changes with style=all_docs.  2) revs_diff on the target
database.
>>>>>>> 3) get each revision from source with revs=true.  4) bulk post
with
>>>>>>> new_edits=false.
>>>>>>> 
>>>>>>> Same thing.  Except now I can kind of make sense of whats going
on.
>>>>>>> Sucking the data out of the source is no problem.  Diffing the
revs
>>>>>>> against the target is no problem.  Posting the docs is THE problem.
 Since
>>>>>>> the database is clean, thousands of docs are being thrown at
couch at once
>>>>>>> to build up the revision trees.  Couch is just taking forever
in finishing
>>>>>>> the job.  It doesn't matter if I bulk post the docs or post them
>>>>>>> individually, couch sucks 100% of my cpu every time and takes
forever to
>>>>>>> finish. (I actually never let it finish).
>>>>>>> 
>>>>>>> So that is is the story. Here are my questions.
>>>>>>> 
>>>>>>> 1) Has anyone else stepped on this mine?  If so, could I get
pointed
>>>>>>> towards some workarounds?  I don't think it is right to make
the assumption
>>>>>>> that users of couchdb will never have databases with huge conflict
sausages
>>>>>>> like this. So simply saying manage your conflicts won't help.
>>>>>>> 
>>>>>>> 2) Lets say I did manage my conflicts.  I still have the
>>>>>>> _deleted_conflicts sausage.  I know that _deleted and _deleted_docs
must be
>>>>>>> replicated to maintain consistency across the cluster.  If the
replicator
>>>>>>> throws up when these huge sausages come through, how is the data
ever going
>>>>>>> to replicate?  Is there a trade secret I don't know about?
>>>>>>> 
>>>>>>> 3) Is there any limit on the resources that CouchDB is allowed
to consume?
>>>>>>> I can get that we run into these cases where theres tons of data
to move
>>>>>>> and its just going to take a hell of a long time.  But I don't
get why its
>>>>>>> permissible for CouchDB to eat all my cpu.  The whole server
should never
>>>>>>> grind to a halt because its moving lots of data.  I feel like
it should be
>>>>>>> like the little train who could.  Just chug along slow and steady
until it
>>>>>>> crests the hill.
>>>>>>> 
>>>>>>> I would really like to reply on the erlang replicator, but I
can't.  At
>>>>>>> least with the replicator I wrote I have a chance with throttling
the posts
>>>>>>> so CouchDB doesn't render my server useless.
>>>>>>> 
>>>>>>> Sorry for wrapping more questions into those questions.  I'm
pretty tired,
>>>>>>> stumped, and have machines in production crumbling.
>>>>>>> 
>>>>>>> Best,
>>>>>>> Stephen
>>>> 
>> 


Mime
View raw message