You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
It would be a performance optimization to combine the two. This also would make the log_append_latency statistic truly account for the number of write calls.
The text was updated successfully, but these errors were encountered:
@spolitov the statistics log_append_latency and log_bytes_logged only account for the what I see as second write, which is probably the actual WAL data, and the first one being something like a header. It would be good to have these reflect what actually is performed at the OS level.
…ices in WritableLogSegment::WriteEntryBatch
Summary:
Currently, we call writable_file_->Append twice in WritableLogSegment::WriteEntryBatch.
It could be replaced with writable_file_->AppendSlices to minimize the amount of system calls.
Test Plan: Jenkins
Reviewers: rthallam
Reviewed By: rthallam
Subscribers: kannan, fhoogland, ybase, bogdan
Differential Revision: https://phabricator.dev.yugabyte.com/D14963
Description
Currently for a WAL write the header belonging to the WAL data is written completely separately from the actual data, leading to two system calls.
This happens in WriteableLogSegment::WriteEntryBatch: https://sourcegraph.com/github.com/yugabyte/[email protected]/-/blob/src/yb/consensus/log_util.cc?L863:28
(line 881+884)
It would be a performance optimization to combine the two. This also would make the log_append_latency statistic truly account for the number of write calls.
The text was updated successfully, but these errors were encountered: