Bulk containing updates on the same document ~30x slower on es 5.0.0

[TrustNoOne]

Environment:
Elasticsearch 5.0.0: vanilla, single node, fresh install, no custom configuration or plugins

JVM version: 1.8.0_74

OS version: tried on mac os and linux

Description of the problem including expected versus actual behavior:
The same exact bulk containing 1 index request and 1000 updates of the same document takes ~500-700ms (on my hardware) on es 2.4 and ~12-15s on es 5.0.0. I tried on a linux machine and get the same result.
Tried transport client as well, no difference.

Am I missing something?

Steps to reproduce:

1. create a simple mapping with a single field (I tried long and keyword, same thing) on both es 2.4 and 5.0
2. create a bulk containing one index request for a document and 1000 different updates for the same document
3. run the bulk query on es 2.4 and 5.0.

Bulk request Example
message is a long field in this example, but I've tried keyword/string not analyzed as well

{ "index" : { "_index" : "twitter", "_type" : "tweet", "_id" : "1" } }
{ "message" : 0 }
{ "update" : {"_id" : "1", "_type" : "tweet", "_index" : "twitter"} }
{ "doc" : {"message" : 1} }
{ "update" : {"_id" : "1", "_type" : "tweet", "_index" : "twitter"} }
{ "doc" : {"message" : 2} }
{ "update" : {"_id" : "1", "_type" : "tweet", "_index" : "twitter"} }
{ "doc" : {"message" : 3} }
{ "update" : {"_id" : "1", "_type" : "tweet", "_index" : "twitter"} }
{ "doc" : {"message" : 4} }
...
{ "update" : {"_id" : "1", "_type" : "tweet", "_index" : "twitter"} }
{ "doc" : {"message" : 1000} }

[clintongormley]

Hi @TrustNoOne

This is a result of this change: #20102 which removed the read-unrefreshed-changes-from-transaction-log behaviour in favour of performing a refresh and reading from Lucene. This greatly reduced the memory requirements for indexing leaving more for the index buffer, reduced young garbage collection, and had a big impact on indexing performance.

The downside is that frequent updates to the same document are slower, and your benchmark is a worst case scenario. The logic is that very frequent updates (within 1 second) to the same document is an edge case. Usually these updates are interspersed with other changes, which would greatly reduce the number of refreshes that are required.

[TrustNoOne]

I thought this was by design but couldn't find anything, thank you for the quick reply!