This is a break-off ticket from
WT-6796 as we want to understand why delete throughout decreases as some of the remove-only benchmarks progress.
There are few things to consider here:
- As discussed in
WT-6796, the reduced delete operation doesn't seem to correlate with history store file size.
- The delete throughout is decreasing as we keep deleting random documents from the test collections without replacing them.
- The benchmark is specifically testing system performance with long running transactions. From the plot above we can notice that the stat "range of timestamps currently pinned" is growing to 1 hour.
- Deleting records in WT essentially entails adding a stop timestamp the records. The actual record in not removed until the stop time becomes globally visible.
- We need to gather more information about the benchmark. For example, how cursors are being used to delete the data and if there any indexes on the test collection.
The aim of this ticket is to find the root cause behind the delete performance degradation and create tickets for possible improvements.