Testing the Write Performance of CouchDB | skybert.net

Testing the Write Performance of CouchDB


First off, I had to get Siege to send the correct content type header. Once that was out of the way, I could go ahead and do my tests :-)

To conduct the tests, I created a.json file with a typical document I would use in my application:

{
  "action_type_id" : "1",
  "content_id" : "123",
  "type" : "slideshow",
  "created_at" : "2011-05-16 12:42",
  "publication_id" : "3",
  "referrer" : "http://mysite.com/add/slideshow/to/one/of/these/blogs",
  "section_id" : "12",
  "target_content_id" : "132",
  "target_content_type" : "blog",
  "target_section_id" : "22",
  "target_user_id" : "44",
  "title" : "The title of the blog 123",
  "uri" : "/blog/123",
  "user_id" : "12"
}

I then created a.siege file for my beloved load testing tool to use. It contained ony one line:

http://127.0.0.1:5984/mydb POST < mydb_entry_example.json

I then unleashed siege:

$ siege -c 100 -f mydb_entry_example.json.siege

The above gave me (running this locally, so network latency, DNS lookup and so on, is not included here):

Transactions:       76700 hits
Availability:      100.00 %
Elapsed time:      393.84 secs
Data transferred:  6.95 MB
Response time:     0.01 secs
Transaction rate:  194.75 trans/sec
Throughput:        0.02 MB/sec
Concurrency:       2.82
Successful transactions:       76700
Failed transactions:           0
Longest transaction:           0.20
Shortest transaction:          0.00

In general, I found that couchdb could not sustain write load over time when siege runs with much more than 100 concurrent sessions. Cranking the number of concurrent TCP sessions, yielded time outs and long error messages in the couchdb:

[Mon, 16 May 2011 07:30:17 GMT] [error] [<0.20137.2>] Uncaught error in HTTP request: {exit,
{timeout,
{gen_server,call,
[couch_query_servers,

Testing CouchDB bulk update

To test CouchDB's bulk update entry point, you have to create a container document which holds all the documents. This was easily done with these three lines of BASH:

$ echo '{"docs":[' > mydb_entry_example-bulk.json
$ for i in {1..100}; do cat mydb_entry_example.json >> mydb_entry_example-bulk.json ; done
$ echo ']}' >> mydb_entry_example-bulk.json

Remember that the container document "docs", also needs to pass the input test of fields required by your application that you may have fined in your validate_doc_update hook.

I created a.siege file like the one for the normal updates:

http://127.0.0.1:5984/mydb/_bulk_docs/ POST < mydb_entry_example.json

I could then run my bulk update tests, first a simple one usingwget:

$ time wget -o /dev/null \
  -S \
  --header "Content-type:application/json" \
  -O /dev/null \
  --post-file=mydb_entry_example-bulk.json  \
  http://127.0.0.1:5984/mydb/_bulk_docs/

real0m2.272s
user0m0.004s
sys0m0.000s

Interestingly enough, there was no time difference creating 100 documents this way or 1000. The reason why I kept the document at 100, was that otherwise couchdb woud throw 400 error messages when siege was unleashed on it.

Running siege was like before, just with different inut parameters:

$ siege -c 30 -f  mydb_entry_example-bulk.json.siege

Lifting it after a while gave these results:

Transactions:        7328 hits
Availability:      100.00 %
Elapsed time:     1132.70 secs
Data transferred: 59.42 MB
Response time:    1.04 secs
Transaction rate: 6.47 trans/sec
Throughput:       0.05 MB/sec
Concurrency:      6.74
Successful transactions:    7328
Failed transactions:        0
Longest transaction:        1.76

These wee tests shows taht using bulk updates improves the CouchDB wirte performacne and (locally on my system) suggests something like100 * 6.47 = 647 writes per second, which isn't that bad :-) As hinted at above, these numbers would probably have been better if siege didn't choke on documents of a certain size (I initially ran it with 1000 document big.json).


Licensed under CC BY Creative Commons License ~ gmail torstein.k.johansen @ gmail ~ twitter @torsteinkrause ~