feat: Add auto scale writer support #11702
Closed
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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
✅ Deploy Preview for meta-velox canceled.
|
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 1, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
61fd2ce to
6e0df03
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 1, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
6e0df03 to
bbe7480
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 1, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
bbe7480 to
d3a8e5f
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 1, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
d3a8e5f to
d2c82c3
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 2, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
d2c82c3 to
2274648
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 2, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
2274648 to
2be3e70
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 2, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
2be3e70 to
4760146
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 2, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
4760146 to
f47d919
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
210a913 to
b36708b
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 2, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
b36708b to
bf6cac0
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
changed the title
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 3, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Reviewed By: oerling Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
bf6cac0 to
141a3cd
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
arhimondr
approved these changes
Dec 4, 2024
zation99
approved these changes
Dec 4, 2024
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 4, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Reviewed By: arhimondr, oerling, zation99 Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
141a3cd to
258e9a3
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 4, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Reviewed By: arhimondr, oerling, zation99 Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
258e9a3 to
7148d31
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
xiaoxmeng
added a commit
to xiaoxmeng/velox
that referenced
this pull request
Dec 4, 2024
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Reviewed By: arhimondr, oerling, zation99 Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
7148d31 to
b7197a0
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
Summary: This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Reviewed By: arhimondr, oerling, zation99 Differential Revision: D66380785
xiaoxmeng
force-pushed
the
export-D66380785
branch
from
b7197a0 to
30706ee
Compare
|
This pull request was exported from Phabricator. Differential Revision: D66380785 |
|
This pull request has been merged in e40259f. |
zuyu
reviewed
Dec 5, 2024
| LocalPartitionNode( | ||
| const PlanNodeId& id, | ||
| Type type, | ||
| bool scaleWriter, |
There was a problem hiding this comment.
@xiaoxmeng how about putting this arg in the end, w/ a default value false?
It seems this feature is disabled by default.
athmaja-n
pushed a commit
to athmaja-n/velox
that referenced
this pull request
Jan 10, 2025
Summary: Pull Request resolved: facebookincubator#11702 This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions. We add two customized local partition operators: ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing if the exchange queue has >50% memory buffering until scale to all the table writer threads; ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling the busy logical table partition by assigning more table writer threads. Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning scheme, and the coordinator needs to configure the query plan accordingly. Reviewed By: arhimondr, oerling, zation99 Differential Revision: D66380785 fbshipit-source-id: 87a414516cc5011a26309db104c6d133c179e959
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.
Summary:
This change adds local scale writer partition support to improve memory efficiency in case of a large number of partitions.
We add two customized local partition operators:
ScaleWriterLocalPartition for non-partitioned table write. It starts with single table writer thread and scale the writer processing
if the exchange queue has >50% memory buffering until scale to all the table writer threads;
ScaleWriterPartitioningLocalPartition for partitioned table writer. It starts with assigning a single table writer thread to each logical
table partition. Multiple physical table partitions could be mapped to a single logical partition based on the partition keys of the
written table. Similar, if the exchange queue has > 50% memory buffering, we leverage the skewed partition balancer by scaling
the busy logical table partition by assigning more table writer threads.
Meta internal shadow results show this could prevent query write OOM pattern, reduce the peak memory usage which benefits the
resource usage accounting which takes into account of accumulated memory usage, it also reduces >2x of written files
The followup is to investigate the more reliable rebalance signal such as consumer/producer queuing delay in the exchange
queue. To complete this feature, we need a Prestissimo change to setup scale writer local partition based on arbitrary partitioning
scheme, and the coordinator needs to configure the query plan accordingly.
Differential Revision: D66380785