feat: Add encodedVectorCopy #12588
Closed
feat: Add encodedVectorCopy #12588
Conversation
This file contains hidden or 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
facebook-github-bot
added
the
CLA Signed
✅ Deploy Preview for meta-velox canceled.
|
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70867237 |
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70867237 |
1 similar comment
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70867237 |
Yuhta
added a commit
to Yuhta/velox
that referenced
this pull request
Mar 10, 2025
Summary: Pull Request resolved: facebookincubator#12588 Implement `encodedVectorCopy`, a generic vector copy utility that preserves encodings for memory saving purpose. ## Encoding Preservation There are mainly 2 use cases for this new function. One is to merge multiple encoded vectors (`source`s) into one large encoded vector (`target`); the other is to update specific rows (`source`) in a large vector (`target`), while keeping the encodings. Both use cases requires us to keep the encoding on `target`, so it is decided as the behavior of this function. There are some exceptions to this rule: - We merge multiple adjacent layers of dictionary and constant wrappers into one. - When `target` is constant, we convert it to dictionary to allow different values in `source`. - When `target` is flat ROW, MAP, or ARRAY, and `source` is constant or dictionary encoded, the result will be dictionary encoded, to avoid flattening the child vectors. Once the target becomes dictionary, it can stay that way and we can keep adding new content to it while keeping the encoding, this is a typical use case for encoding preserved merging. ## Inner Vector Compaction Other than encoding, we also pay special attention to avoid holding on memory that is no longer needed. This is especially important for the merging use case, as the `target` gets updated, majority rows of its inner vectors will be dereferenced and no longer used. There are 2 cases where we need to take care of this. The first is for dictionary encoding, some rows in the alphabet (base/value) vector become no longer referenced by the indices, so we should recycle them. This is done properly that when we translate the copy ranges on dictionary indices to the copy ranges on alphabet, we overwrite the unused rows in alphabet using the new alphabet rows from source. This way we efficiently reuse the memory in alphabet without both reallocation and memory leaking. The second case is for `ARRAY`/`MAP`, the elements/keys/values vector can have rows that are no longer referenced from the parent. This is a little harder to solve than in the dictionary case, since the nested rows need to be contiguous for one parent row (offset/size pair), which means we cannot move them around easily. Our approach is to allow some unused nested rows, but keep track of the percentage of them, and once they exceed certain threshold (50% in the implementation), we make a new copy of the nested vector and copy only the used rows over. This allows us to reuse the nested rows to a certain degree while keep some bounds on the memory usage. Differential Revision: D70867237
Yuhta
force-pushed
the
export-D70867237
branch
2 times, most recently
from
42f9a6a to
ad7d712
Compare
Yuhta
added a commit
to Yuhta/velox
that referenced
this pull request
Mar 11, 2025
Summary: Implement `encodedVectorCopy`, a generic vector copy utility that preserves encodings for memory saving purpose. ## Encoding Preservation There are mainly 2 use cases for this new function. One is to merge multiple encoded vectors (`source`s) into one large encoded vector (`target`); the other is to update specific rows (`source`) in a large vector (`target`), while keeping the encodings. Both use cases requires us to keep the encoding on `target`, so it is decided as the behavior of this function. There are some exceptions to this rule: - We merge multiple adjacent layers of dictionary and constant wrappers into one. - When `target` is constant, we convert it to dictionary to allow different values in `source`. - When `target` is flat ROW, MAP, or ARRAY, and `source` is constant or dictionary encoded, the result will be dictionary encoded, to avoid flattening the child vectors. Once the target becomes dictionary, it can stay that way and we can keep adding new content to it while keeping the encoding, this is a typical use case for encoding preserved merging. ## Inner Vector Compaction Other than encoding, we also pay special attention to avoid holding on memory that is no longer needed. This is especially important for the merging use case, as the `target` gets updated, majority rows of its inner vectors will be dereferenced and no longer used. There are 2 cases where we need to take care of this. The first is for dictionary encoding, some rows in the alphabet (base/value) vector become no longer referenced by the indices, so we should recycle them. This is done properly that when we translate the copy ranges on dictionary indices to the copy ranges on alphabet, we overwrite the unused rows in alphabet using the new alphabet rows from source. This way we efficiently reuse the memory in alphabet without both reallocation and memory leaking. The second case is for `ARRAY`/`MAP`, the elements/keys/values vector can have rows that are no longer referenced from the parent. This is a little harder to solve than in the dictionary case, since the nested rows need to be contiguous for one parent row (offset/size pair), which means we cannot move them around easily. Our approach is to allow some unused nested rows, but keep track of the percentage of them, and once they exceed certain threshold (50% in the implementation), we make a new copy of the nested vector and copy only the used rows over. This allows us to reuse the nested rows to a certain degree while keep some bounds on the memory usage. Reviewed By: mbasmanova Differential Revision: D70867237
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70867237 |
Summary: Implement `encodedVectorCopy`, a generic vector copy utility that preserves encodings for memory saving purpose. ## Encoding Preservation There are mainly 2 use cases for this new function. One is to merge multiple encoded vectors (`source`s) into one large encoded vector (`target`); the other is to update specific rows (`source`) in a large vector (`target`), while keeping the encodings. Both use cases requires us to keep the encoding on `target`, so it is decided as the behavior of this function. There are some exceptions to this rule: - We merge multiple adjacent layers of dictionary and constant wrappers into one. - When `target` is constant, we convert it to dictionary to allow different values in `source`. - When `target` is flat ROW, MAP, or ARRAY, and `source` is constant or dictionary encoded, the result will be dictionary encoded, to avoid flattening the child vectors. Once the target becomes dictionary, it can stay that way and we can keep adding new content to it while keeping the encoding, this is a typical use case for encoding preserved merging. ## Inner Vector Compaction Other than encoding, we also pay special attention to avoid holding on memory that is no longer needed. This is especially important for the merging use case, as the `target` gets updated, majority rows of its inner vectors will be dereferenced and no longer used. There are 2 cases where we need to take care of this. The first is for dictionary encoding, some rows in the alphabet (base/value) vector become no longer referenced by the indices, so we should recycle them. This is done properly that when we translate the copy ranges on dictionary indices to the copy ranges on alphabet, we overwrite the unused rows in alphabet using the new alphabet rows from source. This way we efficiently reuse the memory in alphabet without both reallocation and memory leaking. The second case is for `ARRAY`/`MAP`, the elements/keys/values vector can have rows that are no longer referenced from the parent. This is a little harder to solve than in the dictionary case, since the nested rows need to be contiguous for one parent row (offset/size pair), which means we cannot move them around easily. Our approach is to allow some unused nested rows, but keep track of the percentage of them, and once they exceed certain threshold (50% in the implementation), we make a new copy of the nested vector and copy only the used rows over. This allows us to reuse the nested rows to a certain degree while keep some bounds on the memory usage. Reviewed By: mbasmanova Differential Revision: D70867237
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D70867237 |
Contributor
|
This pull request has been merged in 7d05ec7. |
|
Conbench analyzed the 1 benchmark run on commit There were no benchmark performance regressions. 🎉 The full Conbench report has more details. |
jinchengchenghh
pushed a commit
to jinchengchenghh/velox
that referenced
this pull request
Apr 4, 2025
Summary: Pull Request resolved: facebookincubator#12588 Implement `encodedVectorCopy`, a generic vector copy utility that preserves encodings for memory saving purpose. ## Encoding Preservation There are mainly 2 use cases for this new function. One is to merge multiple encoded vectors (`source`s) into one large encoded vector (`target`); the other is to update specific rows (`source`) in a large vector (`target`), while keeping the encodings. Both use cases requires us to keep the encoding on `target`, so it is decided as the behavior of this function. There are some exceptions to this rule: - We merge multiple adjacent layers of dictionary and constant wrappers into one. - When the values type size in dictionary is no larger than the index type, we flatten the vector to save memory. - When `target` is constant, we convert it to dictionary to allow different values in `source`. - When `target` is flat ROW, MAP, or ARRAY, and `source` is constant or dictionary encoded, the result will be dictionary encoded, to avoid flattening the child vectors. Once the target becomes dictionary, it can stay that way and we can keep adding new content to it while keeping the encoding, this is a typical use case for encoding preserved merging. ## Inner Vector Compaction Other than encoding, we also pay special attention to avoid holding on memory that is no longer needed. This is especially important for the merging use case, as the `target` gets updated, majority rows of its inner vectors will be dereferenced and no longer used. There are 2 cases where we need to take care of this. The first is for dictionary encoding, some rows in the alphabet (base/value) vector become no longer referenced by the indices, so we should recycle them. This is done properly that when we translate the copy ranges on dictionary indices to the copy ranges on alphabet, we overwrite the unused rows in alphabet using the new alphabet rows from source. This way we efficiently reuse the memory in alphabet without both reallocation and memory leaking. The second case is for `ARRAY`/`MAP`, the elements/keys/values vector can have rows that are no longer referenced from the parent. This is a little harder to solve than in the dictionary case, since the nested rows need to be contiguous for one parent row (offset/size pair), which means we cannot move them around easily. Our approach is to allow some unused nested rows, but keep track of the percentage of them, and once they exceed certain threshold (50% by default and configurable), we make a new copy of the nested vector and copy only the used rows over. This allows us to reuse the nested rows to a certain degree while keep some bounds on the memory usage. Reviewed By: mbasmanova Differential Revision: D70867237 fbshipit-source-id: 0cddd37fd7188d89ea541fb89324aa9a10745415
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Summary:
Implement
encodedVectorCopy, a generic vector copy utility that preservesencodings for memory saving purpose.
Encoding Preservation
There are mainly 2 use cases for this new function. One is to merge multiple
encoded vectors (
sources) into one large encoded vector (target); the otheris to update specific rows (
source) in a large vector (target), whilekeeping the encodings. Both use cases requires us to keep the encoding on
target, so it is decided as the behavior of this function.There are some exceptions to this rule:
one.
targetis constant, we convert it to dictionary to allow differentvalues in
source.targetis flat ROW, MAP, or ARRAY, andsourceis constant ordictionary encoded, the result will be dictionary encoded, to avoid flattening
the child vectors. Once the target becomes dictionary, it can stay that way
and we can keep adding new content to it while keeping the encoding, this is a
typical use case for encoding preserved merging.
Inner Vector Compaction
Other than encoding, we also pay special attention to avoid holding on memory
that is no longer needed. This is especially important for the merging use
case, as the
targetgets updated, majority rows of its inner vectors will bedereferenced and no longer used. There are 2 cases where we need to take care
of this.
The first is for dictionary encoding, some rows in the alphabet (base/value)
vector become no longer referenced by the indices, so we should recycle them.
This is done properly that when we translate the copy ranges on dictionary
indices to the copy ranges on alphabet, we overwrite the unused rows in alphabet
using the new alphabet rows from source. This way we efficiently reuse the
memory in alphabet without both reallocation and memory leaking.
The second case is for
ARRAY/MAP, the elements/keys/values vector can haverows that are no longer referenced from the parent. This is a little harder to
solve than in the dictionary case, since the nested rows need to be contiguous
for one parent row (offset/size pair), which means we cannot move them around
easily. Our approach is to allow some unused nested rows, but keep track of the
percentage of them, and once they exceed certain threshold (50% in the
implementation), we make a new copy of the nested vector and copy only the used
rows over. This allows us to reuse the nested rows to a certain degree while
keep some bounds on the memory usage.
Differential Revision: D70867237