(#3015) - Issues GET requests on pull replication in parallel for each b... #3174
Conversation
…h batch. In node.js using default settings (5 connections in the pool + 100 docs per batch) over a HTTP connection to a local PouchDB server perf increased by 8x over previous perf. When replicating down the 0.6 gig NPM CouchDB database over HTTPS using CouchDB locally it took CouchDB about 0.69 seconds/100 records. Without this change it took PouchDB around 30 seconds/100 records. With this change on default settings we reduced time to 4 seconds/100 records. By changing the number of connections to 15 and the batch size to 1000 we got perf down to 0.9 seconds/100 records. So perf got down to only 30% worse than CouchDB. I strongly suspect that when the node.js connection harvesting bug is fixed (so we don't lose our existing connections between batches) and with a little more tweaking (it's clear that we are a bit more 'chatty' than CouchDB's replication client) we should be able to match CouchDB's perf.
|
+1, very pleased. Looks like this closes #2940 as well! |
|
@daleharvey & @nolanlawson - YEAH!!!! Although this is going to make #3185 more urgent as anyone using node.js with this change is going to see a lot of connection failures due to how node 0.10 handles connections. BTW, should I also follow this pattern? That is, submit my PR (with ugly history) to Github, let Travis do it's thing, use the patch trick but instead of committing the patch directly to master submit the patch as a new PR? Just checking because I'd really like to not do more damage. :) |
|
@yaronyg if you push your stuff as a branch to master then it will run the full test You can commit with ugly history then squash when you have a green r+'d PR, the patch thing wont squash the commits so be careful with it |
|
Now I'm completely confused but i figure it's my job to figure out Git, not yours. On my next commit I'll just have to fight through this. |
|
Also we are not strict about seeing a green run in Travis for the exact commit that gets squashed/rebased/pushed; as long as it's green at some point, that's fine. |
...atch.
In node.js using default settings (5 connections in the pool + 100 docs per batch) over a HTTP connection to a local PouchDB server perf increased by 8x over previous perf.
When replicating down the 0.6 gig NPM CouchDB database over HTTPS using CouchDB locally it took CouchDB about 0.69 seconds/100 records. Without this change it took PouchDB around 30 seconds/100 records. With this change on default settings we reduced time to 4 seconds/100 records. By changing the number of connections to 15 and the batch size to 1000 we got perf down to 0.9 seconds/100 records. So perf got down to only 30% worse than CouchDB. I strongly suspect that when the node.js connection harvesting bug is fixed (so we don't lose our existing connections between batches) and with a little more tweaking (it's clear that we are a bit more 'chatty' than CouchDB's replication client) we should be able to match CouchDB's perf.