Do not pre-allocate memory when init vector stream #11211
Conversation
facebook-github-bot
added
the
CLA Signed
✅ Deploy Preview for meta-velox canceled.
|
|
@tanjialiang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
tanjialiang
force-pushed
the
po_fix
branch
5 times, most recently
from
8773f2a
to
e4ec3be
Compare
tanjialiang
changed the title
|
@tanjialiang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
There was a problem hiding this comment.
@tanjialiang nice catch. Maybe shadow in batch production for 1hr duration to see if there is any regression. Please help @kevinwilfong take a look. Thanks!
| @@ -350,6 +350,9 @@ void ByteOutputStream::extend(int32_t bytes) { | |||
| ranges_.emplace_back(); | |||
| current_ = &ranges_.back(); | |||
| lastRangeEnd_ = 0; | |||
| if (bytes == 0) { | |||
There was a problem hiding this comment.
Shall we check this early right after the check?
There was a problem hiding this comment.
Not really. We need above code to initialize the above members.
|
@tanjialiang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@tanjialiang merged this pull request in daeff59. |
|
Conbench analyzed the 1 benchmark run on commit There were no benchmark performance regressions. 🎉 The full Conbench report has more details. |
When creating and initializing vector streams we pre-allocate memory for all rows. This will cause unnecessary memory waste when the actual serialized row have many nulls. The null rows will not be serialized as payload, hence make use of the pre-allocated memory. Remove the pre-allocation to increase memory efficiency.
Memory usage of presto serializer decreased by up to 10x for some meta internal data.