npm is mostly attachments and I haven't seen any issues so far. I wish there was a better way to replicate attachments atomically for a single revision but if there is, I don't know about it.
It's probably a huge JSON operation and it sucks, but I don't have to parse it in node.js, I just pipe() the body right along. -Mikeal On Sep 14, 2011, at September 14, 20118:42 AM, Adam Kocoloski wrote: > Hi Mikeal, I just took a quick peek at your code. It looks like you handle > attachments by inlining all of them into the JSON representation of the > document. Does that ever cause problems when dealing with the ~100 MB > attachments in the npm repo? > > I've certainly seen my fair share of problems with attachment replication in > CouchDB 1.0.x. I have a sneaking suspicion that there are latent bugs related > to incorrect determinations of Content-Length under various compression > scenarios. > > Adam > > On Tuesday, September 13, 2011 at 5:08 PM, Mikeal Rogers wrote: > >> My replicator is fairly young so I think calling it "reliable" might be a >> little misleading. >> >> It does less, I don't ever attempt to cache the high watermark (last seq >> written) and start over from there. If the process crashes just start over >> from scratch. This can lead to a delay after restart but I find that it's >> much simpler and more reliable on failure. >> >> It's also simpler because it doesn't have to content with being an http >> client and a client of the internal couchdb erlang API. It just proxies >> requests from one couch to another. >> >> While I'm sure there are bugs that I haven't found yet in it, I can say that >> it replicates the npm repository quite well and I'm using it in production. >> >> -Mikeal >> >> On Sep 13, 2011, at September 13, 201111:44 AM, Max Ogden wrote: >> >>> Hi Chris, >>> >>> From what I understand the current state of the replicator (as of 1.1) is >>> that for certain types of collections of documents it can be somewhat >>> fragile. In the case of the node.js package repository, http://npmjs.org, >>> there are many relatively large (~100MB) documents that would sometimes >>> throw errors or timeout during replication and crash the replicator, at >>> which point the replicator would restart and attempt to pick up where it >>> left off. I am not an expert in the internals of the replicator but >>> apparently the cumulative time required for the replicator to repeatedly >>> crash and then subsequently relocate itself in _changes feed in the case of >>> replicating the node package manager was making the built in couch >>> replicator unusable for the task. >>> >>> Two solutions exist that I know of. There is a new replicator in trunk (not >>> to be confused with the _replicator db from 1.1 -- it is still using the old >>> replicator algorithms) and there is also a more reliable replicator written >>> in node.js https://github.com/mikeal/replicate that was was written >>> specifically to replicate the node package repository between hosting >>> providers. >>> >>> Additionally it may be useful if you could describe the 'fingerprint' of >>> your documents a bit. How many documents are in the failing databases? are >>> the documents large or small? do they have many attachments? how large is >>> your _changes feed? >>> >>> Cheers, >>> >>> Max >>> >>> On Tue, Sep 13, 2011 at 11:22 AM, Chris Stockton >>> <[email protected] (mailto:[email protected])>wrote: >>> >>>> Hello, >>>> >>>> We now have about 150 dbs that are refusing to replicate with random >>>> crashes, which provide really zero debug information. The error is db >>>> not found, but I know its available. Does anyone know how can I >>>> trouble shoot this? Do we just have to many databases replicating for >>>> couchdb to handle? 4000 is a small number for the massive hardware >>>> these are running on. >>>> >>>> -Chris > >
