Possible Bulk Update Optimizations

[imotov]

We had several reports of slow bulk updates in 5.0 and above. Some of them are summarized in #23792. The main issue that led to significant slow down seems to the refresh that is now performed before each update operation. Basically, in order to perform a bulk requests with 1000 updates on the same record we will have to perform refresh 1000 times. This problem can be avoided if we could combine all update operations on the same record within a shard batch into a single update operation, this way we could get away with performing refresh only once.

For example, if we have the following bulk request:

update rec A
update rec A
update rec B
update rec A

it currently translates into

refresh // possibly if A was updated in the previous bulk and wasn't refreshed yet
get A
update A
index A

refresh // always
get A
update A
index A

// no refresh - because B wasn't modified in this bulk yet
get B
update B
index B

refresh // always
get A
update A
index A

If we combine all updates together we could transform it

refresh
get A
update A
update A
update A
index A
get B
update B
index B

We would still have an issue if we have a mix of index or delete operations with update operations on the same record in the same bulk request, but, perhaps we can either fall back to the current slow way of doing things or optimize them only as much as we can, for example, we can ignore all updates before index operation and all updates and index operation before delete, etc.

[jpountz]

@imotov in it currently translates into, I think we would not refresh before getting B since it is necessarily in the index already due to the refresh that we did for A?

[imotov]

@jpountz you are right, thanks a lot for the clarification. I have updated the issue.

[s1monw]

We spoke about this in fixit-friday. While this optimization is doable on the client side (for now) we would like to first look into a more generic optimization that would hold a secondary (engine-private) index reader that is used internally that would not pay the price of refreshing caches, load global ords, rebuild parent-child relationships etc. It would also mean that we can honor the refresh semantics if a realtime get is used. I will keep this issue open for now and ope a new one to explore the opportunities.

[s1monw]

@imotov I got the index reader change in I was talking about. While batching up those kind of updates might be a two sided sword I am still interested in seeing how it would look like code wise, would it be as simple as sorting the updates by ID and use insertion order as a secondary sort and then run them all consecutively? I think it can be as simple as checking if the previous ID is the same and don't fetch the document again? I really wonder if we can have a prototype that does the sorting in the test as a start?

[imotov]

@s1monw yes, I think this rearrangement shouldn't cause any differences in the final result of the bulk operation (at least on the level we can have any guarantee about). We will have to have a special handling for deletes (clean buffer), but otherwise that should optimize one of the issues.

We will still have an issue of indexing the same record over and over again and sending multiple copies of the same record to replicas, but I am fully sold on progress over perfection here :)

[jasontedor]

We will still have an issue of indexing the same record over and over again and sending multiple copies of the same record to replicas, but I am fully sold on progress over perfection here :)

Can you clarify what you mean here? We have to send these operations to maintain the semantics of the translog as an operation log and for semantics of sequence IDs.

[imotov]

@jasontedor yes if we reindex the same record again and again on the primary shard, we have to send it to replicas as many times. I was just thinking, maybe, if we can index it on the primary only after all sequential operations on the same record are performed we will need to send only one copy to replicas. But as I said, we don't need to bundle that in, it is a somewhat separate optimization.

[jasontedor]

@imotov For a shard-level bulk request, we still send a single bulk request to the replicas containing the results of all the indexing operations on the primary. That is, a shard bulk request is a single replication operation.

[imotov]

@jasontedor yes, the question here is, if a bulk request contains 100 script updates on the same record do we need to send 100 slightly different copies of this record to replicas or we can just send the final version and ignore all intermediate states on both primary and replicas?