Skip per-document stored field reads on sorted indices when no stored fields are present

[fcofdez]

Avoid unnecessary seeks and reads on the stored fields file during
flush in SortingStoredFieldsConsumer for segments that do not contain
stored fields.

Closes #15713

[msfroh]

I'm wondering if you could just skip the flush logic altogether if there are no stored fields in the segment? Then you might be able to avoid the startDocument() and finishDocument() calls for each doc.

Though looking at the behavior in IndexingChain that always calls startStoredFields, and the logic in StoredFieldsConsumer, it looks like we do require that startDocument() and finishDocument() get called for each doc. That's a bummer.

[fcofdez]

Though looking at the behavior in IndexingChain that always calls startStoredFields, and the logic in StoredFieldsConsumer, it looks like we do require that startDocument() and finishDocument() get called for each doc. That's a bummer.

Yes, I think that's a contract that we have to keep. Alternatively we could use a different approach and just write a flag in the stored fields metadata where we track that there are no stored fields, but we can still respond calls to the #document methods. But that seems like a deeper change than this and the remaining overhead should be low compared to all the seeks + reads that we do now.

[romseygeek]

Looks good! I have one question about the test. Also, can you add a CHANGES entry in the 10.5 section?

[romseygeek]

Does the test write to any files other than the StoredFields file? I worry that relying on the filename might be a bit fragile.

[fcofdez]

It also writes into a metadata file and into the fields index files. I understand you concern, but I couldn't find another way of asserting that we don't read n times, do you have other ideas?

[romseygeek]

It's not ideal, but we're not doing Codec randomization here so I think we should be OK.

[romseygeek]

This optimization only works if all docs in the segment have no stored fields, right? Which I think is probably fine, as that's a fairly common case.

[fcofdez]

Yes, that's right 👍

[romseygeek]

Thanks @fcofdez!

[fcofdez]

Thanks for the review @romseygeek!