[
https://issues.apache.org/jira/browse/LUCENE-5591?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13984473#comment-13984473
]
Michael McCandless commented on LUCENE-5591:
--------------------------------------------
Looks great Shai, thanks!
Is avgUpdateSize supposed to be "bytes per doc"? If so, instead of
bitsPerValue, shouldn't we return bitsPerValue/8, maybe rounded up to the
nearest byte? Should we rename the method ... maybe ramBytesPerDoc or
something?
Shouldn't BinaryDocValuesFieldUpdates.avgUpdateSize also include the
docs/offsets/lengths RAM used too?
Separately, I noticed BinaryDocValuesFieldUpdates's add method is doing a
BytesRef.append of each added value ... isn't this slowish (O(N^2) where N =
number of docs that have been updated)? BytesRef.append doesn't use
ArrayUtil.grow to size the array on overflow...
> ReaderAndUpdates should create a proper IOContext when writing DV updates
> -------------------------------------------------------------------------
>
> Key: LUCENE-5591
> URL: https://issues.apache.org/jira/browse/LUCENE-5591
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Shai Erera
> Attachments: LUCENE-5591.patch
>
>
> Today we pass IOContext.DEFAULT. If DV updates are used in conjunction w/
> NRTCachingDirectory, it means the latter will attempt to write the entire DV
> field in its RAMDirectory, which could lead to OOM.
> Would be good if we can build our own FlushInfo, estimating the number of
> bytes we're about to write. I didn't see off hand a quick way to guesstimate
> that - I thought to use the current DV's sizeInBytes as an approximation, but
> I don't see a way to get it, not a direct way at least.
> Maybe we can use the size of the in-memory updates to guesstimate that
> amount? Something like {{sizeOfInMemUpdates * (maxDoc/numUpdatedDocs)}}? Is
> it a too wild guess?
--
This message was sent by Atlassian JIRA
(v6.2#6252)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
