bench: aggregate adding completed ops for reads #721
Conversation
cmd/bbolt/main.go
Outdated
| @@ -1380,10 +1380,10 @@ func (cmd *benchCommand) runReadsSequential(db *bolt.DB, options *BenchOptions, | |||
|
|
|||
| for { | |||
| numReads := int64(0) | |||
| defer func() { results.AddCompletedOps(numReads) }() | |||
There was a problem hiding this comment.
I debated using a defer or manually adding the call at the end of the loop. I ended up doing it with the defer, as there are some cases in which it would need to be called from inside the if condition if the run had an error and outside, too. So, it feels cleaner. However, I'm unsure if too many defers (one per loop) would create issues with the call stack if there are too many iterations.
There was a problem hiding this comment.
Adding defer inside a for loop isn't a good practice.
How about this?
err := func() error {
defer func() { results.AddCompletedOps(numReads) }()
c := tx.Bucket(benchBucketName).Cursor()
for k, v := c.First(); k != nil; k, v = c.Next() {
numReads++
if v == nil {
return ErrInvalidValue
}
}
return nil
}()
if err != nil {
return err
}
There was a problem hiding this comment.
Note that deferred function won't be called until the surrounding function returns. It means the initial motivation of progress notification is broken. My proposed change above can fix it.
There was a problem hiding this comment.
@ahrtr, I applied the suggestion. PTAL :)
|
/lgtm |
|
Also we can also do similar enhancement in |
Hey @ahrtr, unfortunately (as I mentioned in #720 (comment)), changing the |
|
@ivanvc Please resolve the comment, also read #720 (comment) |
Sorry, @ahrtr, for some reason, I missed the notification and didn't read your comment. I'll address it later today. |
Currently, the completed operations are added to the read benchmarks one by one, and given that each operation is atomic, it impacts the benchmark's performance. Change to update only once per cycle, with the total number of reads. Signed-off-by: Ivan Valdes <ivan@vald.es>
ivanvc
force-pushed
the
fix-read-benchmark-metrics-overhead
branch
from
0e36040 to
43c669d
Compare
Currently, the completed operations are added to the read benchmarks one by one, and given that each operation is atomic, it impacts the benchmark's performance. Change to update only once per cycle, with the total number of reads.
Related to #720