[Datasets] [Operator Fusion - 5/N] Add metrics collection to data layer. #32749
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
clarkzinzow
requested review from
ericl,
scv119,
jjyao,
jianoaix and
c21
as code owners
clarkzinzow
force-pushed
the
datasets/feat/data-layer-metrics
branch
2 times, most recently
from
65e4e9d to
1da6e5e
Compare
clarkzinzow
force-pushed
the
datasets/feat/data-layer-metrics
branch
from
1da6e5e to
7b9c78c
Compare
clarkzinzow
commented
Feb 28, 2023
ericl
reviewed
Feb 28, 2023
| for block in block_iter: | ||
| with stats.iter_format_batch_s.timer() if stats else nullcontext(): | ||
| batch = BlockAccessor.for_block(block).to_batch_format(batch_format) | ||
| acc = BlockAccessor.for_block(block) |
There was a problem hiding this comment.
You may want to record the metrics from within BlockAccessor, e.g., BlockAccessor(block, metrics_sink) to avoid duplication of code recording metrics and make the code look nicer.
ericl
reviewed
Feb 28, 2023
| @@ -28,5 +32,7 @@ def fn( | |||
| # This causes different behavior between filter and other map-like | |||
| # functions. We should revisit and try to get rid of this logic. | |||
| yield builder.build() | |||
| if metrics_collector is not None: | |||
There was a problem hiding this comment.
Shall we always make this non None?
There was a problem hiding this comment.
Sure! Initially this was going to allow the legacy path to not be ported, but we can always collect the metrics and then drop them.
ericl
reviewed
Feb 28, 2023
There was a problem hiding this comment.
We should also add some unit tests for the metrics collector (this is getting large... should we merge metrics util code in a dedicated PR?)
Couple other thoughts on the metrics:
- Should we record only rows copied/sliced/concatenated? It seems more directly related to performance than the number of slices.
- Maybe start with fewer metrics
ericl
added
the
@author-action-required
clarkzinzow
force-pushed
the
datasets/feat/data-layer-metrics
branch
5 times, most recently
from
f9fa7f0 to
e157ad9
Compare
clarkzinzow
force-pushed
the
datasets/feat/data-layer-metrics
branch
from
e157ad9 to
1959282
Compare
8 tasks
clarkzinzow
changed the title
8 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds basic metrics collection to the Datasets data layer, focusing on batching and block building. This allows us to capture metrics around data slicing, concatenation, copying, etc.
This PR is partially stacked on changes in #32744. This PR is an e2e instrumentation successor to #33831.
TODOs
num_copiesisn't incremented for slicing, even thoughSimpleBlocksalways copy on.slice(), even ifcopy=False.Checks
git commit -s) in this PR.scripts/format.shto lint the changes in this PR.