[COUCHDB-1092] Storing documents bodies as raw JSON binaries instead of serialized JSON terms

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Won't Fix
• Component/s: Database Core
• Skill Level: Committers Level (Medium to Hard)
• Assignee: Filipe David Borba Manana
• Reporter: Filipe David Borba Manana
• Created: 15/Mar/11 20:53
• Updated: 02/May/11 14:36
• Resolved: 02/May/11 14:36

Description

Currently we store documents as Erlang serialized (via the term_to_binary/1 BIF) EJSON.
The proposed patch changes the database file format so that instead of storing serialized
EJSON document bodies, it stores raw JSON binaries.

The github branch is at: https://github.com/fdmanana/couchdb/tree/raw_json_docs

Advantages:

• what we write to disk is much smaller - a raw JSON binary can easily get up to 50% smaller
  (at least according to the tests I did)

• when serving documents to a client we no longer need to JSON encode the document body
  read from the disk - this applies to individual document requests, view queries with
  ?include_docs=true, pull and push replications, and possibly other use cases.
  We just grab its body and prepend the _id, _rev and all the necessary metadata fields
  (this is via simple Erlang binary operations)

• we avoid the EJSON term copying between request handlers and the db updater processes,
  between the work queues and the view updater process, between replicator processes, etc

• before sending a document to the JavaScript view server, we no longer need to convert it
  from EJSON to JSON

The changes done to the document write workflow are minimalist - after JSON decoding the
document's JSON into EJSON and removing the metadata top level fields (_id, _rev, etc), it
JSON encodes the resulting EJSON body into a binary - this consumes CPU of course but it
brings 2 advantages:

1) we avoid the EJSON copy between the request process and the database updater process -
for any realistic document size (4kb or more) this can be very expensive, specially
when there are many nested structures (lists inside objects inside lists, etc)

2) before writing anything to the file, we do a term_to_binary([Len, Md5, TheThingToWrite])
and then write the result to the file. A term_to_binary call with a binary as the input
is very fast compared to a term_to_binary call with EJSON as input (or some other nested
structure)

I think both compensate the JSON encoding after the separation of meta data fields and non-meta data fields.

The following relaximation graph, for documents with sizes of 4Kb, shows a significant
performance increase both for writes and reads - especially reads.

http://graphs.mikeal.couchone.com/#/graph/698bf36b6c64dbd19aa2bef63400b94f

I've also made a few tests to see how much the improvement is when querying a view, for the
first time, without ?stale=ok. The size difference of the databases (after compaction) is
also very significant - this change can reduce the size at least 50% in common cases.

The test databases were created in an instance built from that experimental branch.
Then they were replicated into a CouchDB instance built from the current trunk.
At the end both databases were compacted (to fairly compare their final sizes).

The databases contain the following view:
{
"_id": "_design/test",
"language": "javascript",
"views": {
"simple": {
"map": "function(doc)

{ emit(doc.float1, doc.strings[1]); }

"
}
}
}

1. 1. Database with 500 000 docs of 2.5Kb each

Document template is at: https://github.com/fdmanana/couchdb/blob/raw_json_docs/doc_2_5k.json

Sizes (branch vs trunk):

$ du -m couchdb/tmp/lib/disk_json_test.couch
1996 couchdb/tmp/lib/disk_json_test.couch

$ du -m couchdb-trunk/tmp/lib/disk_ejson_test.couch
2693 couchdb-trunk/tmp/lib/disk_ejson_test.couch

Time, from a user's perpective, to build the view index from scratch:

$ time curl http://localhost:5984/disk_json_test/_design/test/_view/simple?limit=1

{"total_rows":500000,"offset":0,"rows":[ {"id":"0000076a-c1ae-4999-b508-c03f4d0620c5","key":null,"value":"wfxuF3N8XEK6"}

]}

real 6m6.740s
user 0m0.016s
sys 0m0.008s

$ time curl http://localhost:5985/disk_ejson_test/_design/test/_view/simple?limit=1

{"total_rows":500000,"offset":0,"rows":[ {"id":"0000076a-c1ae-4999-b508-c03f4d0620c5","key":null,"value":"wfxuF3N8XEK6"}

]}

real 15m41.439s
user 0m0.012s
sys 0m0.012s

1. 1. Database with 100 000 docs of 11Kb each

Document template is at: https://github.com/fdmanana/couchdb/blob/raw_json_docs/doc_11k.json

Sizes (branch vs trunk):

$ du -m couchdb/tmp/lib/disk_json_test_11kb.couch
1185 couchdb/tmp/lib/disk_json_test_11kb.couch

$ du -m couchdb-trunk/tmp/lib/disk_ejson_test_11kb.couch
2202 couchdb-trunk/tmp/lib/disk_ejson_test_11kb.couch

Time, from a user's perpective, to build the view index from scratch:

$ time curl http://localhost:5984/disk_json_test_11kb/_design/test/_view/simple?limit=1

{"total_rows":100000,"offset":0,"rows":[ {"id":"00001511-831c-41ff-9753-02861bff73b3","key":null,"value":"2fQUbzRUax4A"}

]}

real 4m19.306s
user 0m0.008s
sys 0m0.004s

$ time curl http://localhost:5985/disk_ejson_test_11kb/_design/test/_view/simple?limit=1

{"total_rows":100000,"offset":0,"rows":[ {"id":"00001511-831c-41ff-9753-02861bff73b3","key":null,"value":"2fQUbzRUax4A"}

]}

real 18m46.051s
user 0m0.008s
sys 0m0.016s

All in all, I haven't seen yet any disadvantage with this approach. Also, the code changes
don't bring additional complexity. I say the performance and disk space gains it gives are
very positive.

This branch still needs to be polished in a few places. But I think it isn't far from getting mature.

Other experiments that can be done are to store view values as raw JSON binaries as well (instead of EJSON)
and optional compression of the stored JSON binaries (since it's pure text, the compression ratio is very high).
However, I would prefer to do these other 2 suggestions in separate branches/patches - I haven't actually tested
any of them yet, so maybe they not bring significant gains.

Thoughts?