fix: reproducer for the empty snapshot file issue

[praveen-influx]

No issue yet, this is a reproducer test for empty snapshot files being generated under some very strict settings,

• flush interval: 1s
• snapshot size: 1
• force snapshotting mem threshold: 2000 bytes

The good news is, after investigating this code there is no data loss or any possibility of inconsistent query results. However you see these snapshot files with 0 rows in them. Loading them (during restart) does not cause any panics. It loads the file and ignores it as it does not have any pointers to parquet files. It would be good to fix this nevertheless.

[praveen-influx]

This is not really used in the test or the code change, but it was useful to get the parquet files (as get_files method needed db_id and table_id) when debugging. I can remove it, just left it there as it might be handy.

[praveen-influx]

This check is only needed here to fix the issue but I've also made clearing the snapshots further down conditional (whether persist jobs were empty or not). I did wonder if I can work around the problem further up the call stack to avoid snapshotting (calling this function) in the first place but knowing whether query buffer is empty isn't straight forward. If there is a better way to achieve this without scattering checks on a flag that would be desirable.

[hiltontj]

This LGTM, and the comment with explanation is very helpful so thanks for adding that.

I was trying to reason if the same logic would apply to the catalog. That is, if there is nothing in the buffer, then should we ignore persisting the catalog. But I think the answer is no: there could have been catalog only operations in the WAL, e.g., from CreateTable, that did not result in anything in the buffer...

[praveen-influx]

I was trying to reason if the same logic would apply to the catalog. That is, if there is nothing in the buffer, then should we ignore persisting the catalog. But I think the answer is no

There is already a guard to write new catalog file only when catalog has been updated (see here). I think the condition this PR addresses should be mutually exclusive i.e, cannot have both a no-op (written only when there's nothing in wal buffer) and also a catalog op at the same time.