Groovy, that's consensus then. 

Where we can use _db_updates to know which pending jobs are worth running, we 
will. If it's a 404, we'll do something less optimal, start the job itself. 
Using the connection pool as discussed and accounting for and penalising jobs 
that complete very quickly or perform no useful work by applying a rising retry 
interval should mitigate some of that cost. Happily, we discussed those kinds 
of scheduler nuances this week. 


> On 20 Mar 2016, at 14:54, Adam Kocoloski <[email protected]> wrote:
> 
> I’ll never berate anyone for top-posting (or bottom-posting for that matter). 
> I just follow suit with whatever the current thread is doing — in this, very 
> very clearly top-posting ;)
> 
> Thank you for making this distinction clear. Personally I was only ever 
> interested in the first case. Scoping the replicator manager to only learn 
> about _replicator docs on the local cluster through internal APIs is a smart 
> move — I wasn’t suggesting anything different there. I do think we should 
> have an efficient way for a client to learn about the existence of new 
> updates to an arbitrary number of databases on a remote cluster using a 
> single socket.
> 
> Adam
> 
>> On Mar 20, 2016, at 10:45 AM, Robert Samuel Newson <[email protected]> 
>> wrote:
>> 
>> Final note, we've conflated two uses of /_db_updates that I want to be very 
>> clear on;
>> 
>> 1) using /_db_updates to detect active source databases of a replication job.
>> 2) using /_db_updates to hear about new/updated/deleted _replicator 
>> documents.
>> 
>> It was the 2nd case where the unreliability was a concern, since the update 
>> frequency is very low, one expects, for _replicator databases.
>> 
>>> On 20 Mar 2016, at 14:44, Robert Samuel Newson <[email protected]> wrote:
>>> 
>>> (I swear I'll stop soon...)
>>> 
>>> Using /_db_updates as a cheap mechanism to detect activity at the source 
>>> for any database we're interested in is an important optimization. We 
>>> didn't discuss it this past week as we felt that /_db_updates wasn't 
>>> sufficiently reliable. We can save a lot of churn in the scheduler by 
>>> simply not resuming any job unless we have seen an update to the source 
>>> database.
>>> 
>>> B.
>>> 
>>>> On 20 Mar 2016, at 14:36, Robert Samuel Newson <[email protected]> wrote:
>>>> 
>>>> I missed a point in Adam's earlier post.
>>>> 
>>>> The current scheme uses couch_event for runtime changes to _replicator 
>>>> docs but has to read all updates of all _replicator databases at startup. 
>>>> In the steady state it is just receiving couch_event notifications. The 
>>>> /_db_updates option would change that only slightly (we'd read 
>>>> /_db_updates from 0 to find all _replicator databases, rather than reading 
>>>> the changes feed for the node-local 'dbs' database).
>>>> 
>>>> CouchDB itself has a single /_replicator database, of course, but the code 
>>>> will consider any database to be a /_replicator database if the name ends 
>>>> that way. i.e, today, if you made a database called foo/_replicator it 
>>>> would be considered a /_replicator database by the system (and we'd inject 
>>>> the ddoc, etc).
>>>> 
>>>> B.
>>>> 
>>>>> On 20 Mar 2016, at 14:31, Robert Samuel Newson <[email protected]> wrote:
>>>>> 
>>>>> Since I'm typing anyway, and haven't yet been dinged for top-posting, I 
>>>>> wanted to mention one other optimization we had in mind.
>>>>> 
>>>>> Currently each replicator job has its own connection pool. When we 
>>>>> introduce the notion that we can stop and restart jobs, those become 
>>>>> approximately useless. So we will obvious hoist that 'up' to a higher 
>>>>> level and manage connection pools at the manager level.
>>>>> 
>>>>> One optimization that seems obvious from the Cloudant perspective is to 
>>>>> allow reuse of connections to the same destinations even though they are 
>>>>> ostensibly for different domains. That is, a connection to 
>>>>> rnewson.cloudant.com is ultimately a connection to 
>>>>> lbX.jenever.cloudant.com. This connection could just as easily be used 
>>>>> for any other user in the jenever cluster. Thus, if it's idle, we could 
>>>>> borrow that connection rather than create a new one.
>>>>> 
>>>>>> host rnewson.cloudant.com
>>>>> rnewson.cloudant.com is an alias for jenever.cloudant.com.
>>>>> jenever.cloudant.com is an alias for lb2.jenever.cloudant.com.
>>>>> lb2.jenever.cloudant.com has address 5.153.0.207
>>>>> 
>>>>> Rather than add rnewson.cloudant.com > 5.153.0.207 to the pool, we would 
>>>>> add lb2.jenever.cloudant.com -> 5.153.0.207 and resolve 
>>>>> rnewson.cloudant.com to its ultimate CNAME before consulting the pool.
>>>>> 
>>>>> Does this optimization help elsewhere than Cloudant?
>>>>> 
>>>>>> On 20 Mar 2016, at 14:22, Robert Samuel Newson <[email protected]> 
>>>>>> wrote:
>>>>>> 
>>>>>> My point is that we can (and currently do) trigger the replication 
>>>>>> manager on receipt of the database updated event, so it avoids all of 
>>>>>> the other parts of the sequence you describe which could fail.
>>>>>> 
>>>>>> The obvious difference, and I suspect this is what motivates Adam's 
>>>>>> position, is that _db_updates can be called remotely. A solution using 
>>>>>> /_db_updates as its feed can run somewhere else, it wouldn't even need 
>>>>>> to be a couchdb cluster. With the current 2.0 scheme, the _replicator db 
>>>>>> has to live on the nodes performing replication management (and 
>>>>>> therefore it depends on couch_{btree,file} etc). That's a huge incentive 
>>>>>> to go the /_db_updates route and it would serve as a model for others 
>>>>>> like pouchdb that cannot choose to co-locate.
>>>>>> 
>>>>>> One side-benefit we get from using database updated events from the 
>>>>>> _replicator shards, though, is that it helps us determine which node 
>>>>>> will run any particular job. We allocate a job to the lowest live erlang 
>>>>>> node that hosts the document. If we go with /_db_updates, we'll need 
>>>>>> some other scheme. That's not a bad thing (indeed, it could be a very 
>>>>>> good thing), but it would need more thought. While in Seattle we did 
>>>>>> discuss both directions at some length and believe we'd need some form 
>>>>>> of leader election system, the leader would then assign (and rebalance) 
>>>>>> replication jobs across the erlang cluster. I pointed at a 
>>>>>> proof-of-concept implementation of an algorithm I trust that I wrote a 
>>>>>> while back at https://github.com/cloudant/sobhuza as a possible starting 
>>>>>> point.
>>>>>> 
>>>>>> B.
>>>>>> 
>>>>>> P.S. I'm using Mail.app and simply replying where it sticks the cursor 
>>>>>> (at the top), but in other forums I've been berated for top-posting. 
>>>>>> Should I modify my reply style here?
>>>>>> 
>>>>>>> On 19 Mar 2016, at 21:42, Benjamin Bastian <[email protected]> wrote:
>>>>>>> 
>>>>>>> When a shard is updated, it'll trigger a "database updated" event. 
>>>>>>> CouchDB
>>>>>>> will hold those updates in memory for a configurable amount of time in
>>>>>>> order to dedupe updates. It'll then cast lists of updated databases to
>>>>>>> nodes which host the relevant _db_updates shards for further 
>>>>>>> deduplication.
>>>>>>> It's only at that point that the updates are persisted. Only a single
>>>>>>> update needs to reach the _db_updates DB. IIRC, _db_updates triggers up 
>>>>>>> to
>>>>>>> n^3 (assuming the _db_updates DB and the updated DB have the same N), 
>>>>>>> so it
>>>>>>> may be a bit tricky for all of them to fail. You'd need coordinated node
>>>>>>> failure. Perhaps something like datacenter power loss. Another possible
>>>>>>> issue is if all the nodes which host a shard range of the _db_updates DB
>>>>>>> are unreachable by the nodes which host a shard range of any other DB. 
>>>>>>> Even
>>>>>>> if it was momentary, it'd cause messages to be dropped from the 
>>>>>>> _db_updates
>>>>>>> feed.
>>>>>>> 
>>>>>>> For n=3 DBs, it seems like it'd be difficult for all of those things to 
>>>>>>> go
>>>>>>> wrong (except perhaps in the case of power loss or catastrophic network
>>>>>>> failure). For n=1 DBs, you'd simply need to reboot a node soon after an
>>>>>>> update.
>>>>>>> 
>>>>>>>> On Sat, Mar 19, 2016 at 1:31 PM, Adam Kocoloski <[email protected]> 
>>>>>>>> wrote:
>>>>>>>> 
>>>>>>>> Hi Bob, comments inline:
>>>>>>>> 
>>>>>>>>>> On Mar 19, 2016, at 2:36 PM, Robert Samuel Newson 
>>>>>>>>>> <[email protected]>
>>>>>>>>> wrote:
>>>>>>>>> 
>>>>>>>>> Hi,
>>>>>>>>> 
>>>>>>>>> The problem is that _db_updates is not guaranteed to see every update,
>>>>>>>> so I think it falls at the first hurdle.
>>>>>>>> 
>>>>>>>> Do you mean to say that a listener of _db_updates is not guaranteed to 
>>>>>>>> see
>>>>>>>> every updated *database*? I think it would be helpful for the 
>>>>>>>> discussion to
>>>>>>>> describe the scenario in which an updated database permanently fails to
>>>>>>>> show up in the feed. My recollection is that it’s quite byzantine.
>>>>>>>> 
>>>>>>>>> What couch_replicator_manager does in couchdb 2.0 (though not in the
>>>>>>>> version that Cloudant originally contributed) is to us ecouch_event, 
>>>>>>>> notice
>>>>>>>> which are to _replicator shards, and trigger management work from that.
>>>>>>>> 
>>>>>>>> Did you mean to say “couch_event”? I assume so. You’re describing how 
>>>>>>>> the
>>>>>>>> replicator manager discovers new replication jobs, not how the jobs
>>>>>>>> discover new updates to source databases specified by replication jobs.
>>>>>>>> Seems orthogonal to me unless I missed something.
>>>>>>>> 
>>>>>>>>> Some work I'm embarking on, with a few other devs here at Cloudant, is
>>>>>>>> to enhance the replicator manager to not run all jobs at once and it is
>>>>>>>> indeed the plan to have each of those jobs run for a while, kill them 
>>>>>>>> (they
>>>>>>>> checkpoint then close all resources) and reschedule them later. It's 
>>>>>>>> TBD
>>>>>>>> whether we'd always strip feed=continuous from those. We _could_ let 
>>>>>>>> each
>>>>>>>> job run to completion (i.e, caught up to the source db as of the start 
>>>>>>>> of
>>>>>>>> the replication job) but I think we have to be a bit smarter and allow
>>>>>>>> replication jobs that constantly have work to do (i.e, the source db is
>>>>>>>> always busy), to run as they run today, with feed=continuous, unless
>>>>>>>> forcibly ousted by a scheduler due to some configuration concurrency
>>>>>>>> setting.
>>>>>>>> 
>>>>>>>> So I think this is really the crux of the issue. My contention is that
>>>>>>>> permanently occupying a socket for each continuous replication with the
>>>>>>>> same source and mediator is needlessly expensive, and that _db_updates
>>>>>>>> could be an elegant replacement.
>>>>>>>> 
>>>>>>>>> I note  for completeness that the work we're planning explicitly
>>>>>>>> includes "multi database" strategies, you'll hopefully be able to make 
>>>>>>>> a
>>>>>>>> single _replicator doc that represents your entire intention (e.g,
>>>>>>>> "replicate _all_ dbs from server1 to server2”).
>>>>>>>> 
>>>>>>>> Nice! It’ll be good to hear more about that design as it evolves,
>>>>>>>> particularly in aspects like discovery of newly created source 
>>>>>>>> databases
>>>>>>>> and reporting of 403s and other fatal errors.
>>>>>>>> 
>>>>>>>> Adam
> 

Reply via email to