Fix searchable snapshot stats fwd/bwd seeking counters

[tlrx]

This pull request fixes the backward/forward seeking counters in searchable snapshot statistics. These counters use the current file pointer position and the new seeking position to compute the seeking stats, but at the time the stat is computed the current file pointer position is already updated to the new position, making the delta always equals to 0.

Sadly, no tests were able to catch this.

[elasticmachine]

Pinging @elastic/es-distributed (:Distributed/Snapshot/Restore)

[DaveCTurner]

Oops, that'd explain it 😁

Do we need to distinguish lastSeekPosition from lastReadPosition? I think they only differ between a seek and the subsequent read.

[tlrx]

Do we need to distinguish lastSeekPosition from lastReadPosition?

This allows to track seeks independently from reads (like it is done in the new added tests) and track stats for access patterns like read -> seek to A -> seek to B -> read.

If we don't distinguish lastSeekPosition from lastReadPosition then seeking counters are tracking the seeking length from the last read position; access patterns like read -> seek to A -> seek to B -> read won't be correctly tracked in stats. I do agree that it is not the common access pattern for Lucene files, but at least we're tracking seeks correctly. If we're only interested in tracking seeks that are followed by a read operations then maybe we should compute this information as a read position delta in non-contiguous read stats.

[DaveCTurner]

Ok, makes sense. It'll be interesting to see if we do ever seek twice between reads like that.

LGTM

[tlrx]

Thanks David!