Uh oh!
There was an error while loading. Please reload this page.
Add __slots__ to multiple classes for memory optimization - #647
Closed
mykaul wants to merge 9 commits into
Closed
Conversation
mykaul
marked this pull request as draft
January 15, 2026 13:39
- Add __slots__ definition to _Frame class in connection.py - Optimizes memory usage for protocol frame objects - _Frame is created for every protocol message, making this a high-impact optimization - Reduces object overhead from ~300+ bytes to ~40-60 bytes per frame - All existing functionality preserved, comprehensive tests pass Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
mykaulforce-pushed
the
slots
branch
4 times, most recently
from
January 17, 2026 08:28
d1dbc45 to
e029a81CompareAdd __slots__ to _MessageType base class and all protocol message classes to reduce memory overhead. Each message instance saves approximately 280-300 bytes by eliminating the per-instance __dict__. Changes: - _MessageType: Added __slots__ with 'custom_payload' and 'tracing' - _MessageType: Added __init__ to initialize attributes with proper defaults - _DecodableMessageType: Removed duplicate 'custom_payload' from __slots__ - All message subclasses: Added super().__init__() calls for proper initialization Key attributes must be in __slots__ because: - custom_payload: Accessed by encode_message() for ALL message types (line 1127) - tracing: Set on message instances in cluster.py (line 2972) Without these in __slots__, attempting to set them raises AttributeError, causing connection failures with: 'OptionsMessage' object has no attribute 'custom_payload' Message classes covered: - Outgoing (_MessageType): StartupMessage, OptionsMessage, QueryMessage, ExecuteMessage, PrepareMessage, BatchMessage, RegisterMessage, etc. - Incoming (_DecodableMessageType): ResultMessage, EventMessage, AuthenticateMessage, SupportedMessage, etc. Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
mykaulforce-pushed
the
slots
branch
2 times, most recently
from
January 17, 2026 10:56
23cd816 to
c6539c5CompareImplemented __slots__ for ColumnMetadata and IndexMetadata classes in cassandra/metadata.py to reduce per-instance memory overhead. Changes: - ColumnMetadata: Added __slots__ with 6 attributes (table, name, cql_type, is_static, is_reversed, _cass_type) - IndexMetadata: Added __slots__ with 5 attributes (keyspace_name, table_name, name, kind, index_options) - Removed class-level attribute definitions that conflicted with __slots__ declarations - All attribute initialization moved to __init__ methods - Preserved attribute documentation in class docstrings Memory Impact: Each instance saves approximately 280-300 bytes by eliminating the per-instance __dict__. These metadata objects are created frequently when parsing schema information, making this optimization valuable for clusters with many tables and indexes. Testing: - All 648 unit tests pass successfully - Metadata-specific tests: 53 passed, 2 skipped - No functionality changes or regressions Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
Implemented __slots__ for tablet-related classes in cassandra/tablets.py to reduce per-instance memory overhead. Changes: - Tablet class: Added __slots__ with 3 attributes (first_token, last_token, replicas) - Tablets class: Added __slots__ with 2 attributes (_tablets, _lock) - Removed class-level attribute definitions that conflicted with __slots__ - Preserved and enhanced attribute documentation in class docstrings Memory Impact: Each Tablet instance saves approximately 280-300 bytes by eliminating the per-instance __dict__. Tablet objects are created for each token range in ScyllaDB tablets-enabled clusters. In a cluster with many tablets, this optimization significantly reduces the driver's memory footprint. The Tablets class is typically a singleton per cluster, so the memory savings are minimal there, but the change maintains consistency with the optimization pattern and prevents future __dict__ usage. Testing: - All 7 tablet-specific unit tests pass - Verified __dict__ is not created (memory optimization confirmed) - No functionality changes or regressions Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
These attributes are defined in ResponseFuture but never used: - default_timeout: Only used in Session class - _profile_manager: Only used in Session class - _warned_timeout: Never referenced anywhere Removing these attributes before adding __slots__ to ensure clean slot definitions. Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
ResponseFuture is created for every query execution, making it a high-impact target for memory optimization. With __slots__, each instance saves approximately 900 bytes by eliminating __dict__. Memory savings at 1,000 queries/sec: ~904 KB/sec Memory savings at 10,000 queries/sec: ~9 MB/sec Changes: - Added __slots__ tuple with all 30 instance attributes - Removed class-level attribute definitions (conflicts with __slots__) - Initialize all attributes with proper defaults in __init__ - Updated test_repeat_orig_query_after_succesful_reprepare to patch at class level instead of instance level (required for __slots__) Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
Implements __slots__ for Statement, SimpleStatement, BoundStatement, and BatchStatement classes to reduce memory overhead by eliminating per-instance __dict__. Measured memory savings (empirically verified): - Statement: 240 bytes saved (344 → 104 bytes) = 69.8% reduction - SimpleStatement: 232 bytes saved (344 → 112 bytes) = 67.4% reduction - BoundStatement: inherits full parent optimization - BatchStatement: inherits full parent optimization Classes modified: - Statement: Added __slots__ with 9 attributes (retry_policy, consistency_level, fetch_size, keyspace, table, custom_payload, is_idempotent, _serial_consistency_level, _routing_key) - SimpleStatement: Added __slots__ with 1 additional attribute (_query_string) - BoundStatement: Added __slots__ with 3 additional attributes (prepared_statement, values, raw_values) - BatchStatement: Added __slots__ with 4 additional attributes (batch_type, _statements_and_parameters, _session, _is_lwt) All 562 unit tests pass. No behavior changes. Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
mykaul
marked this pull request as ready for review
January 18, 2026 07:26
This commit introduces __slots__ to the Host and HostConnection classes in cassandra/pool.py to reduce memory footprint of frequently instantiated objects in connection pools and cluster metadata. Changes: - Added __slots__ tuple to Host class with 18 actively used attributes - Added __slots__ tuple to HostConnection class with 17 actively used attributes - Consolidated class documentation from individual attribute docstrings to comprehensive class-level docstrings - Added explicit attribute initialization in __init__ methods to set default values for attributes that previously relied on class variables - Removed unused attributes (listen_address, listen_port) from Host class after comprehensive usage analysis showed they were never used - Fixed trailing whitespace issues in modified code The __slots__ implementation prevents dynamic attribute creation and reduces per-instance memory overhead, which is significant for drivers managing large numbers of hosts and connections. All existing functionality is preserved and tests continue to pass. Memory impact: Each Host instance saves ~2 unused slots, and both classes benefit from __slots__ memory layout optimization. Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>
Add __slots__ to frequently instantiated token-related classes to reduce
memory overhead by preventing dynamic attribute creation:
* Token base class with __slots__ = ('value',)
* All Token subclasses (HashToken, Murmur3Token, MD5Token, BytesToken)
with __slots__ = ()
* TokenMap class with comprehensive attribute tuple and consolidated
docstring documentation
These classes are instantiated frequently during cluster metadata parsing
and query planning. The __slots__ optimization prevents Python from
creating __dict__ for each instance, significantly reducing memory usage
especially in large clusters with many tokens.
All existing functionality is preserved and unit tests pass.
Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com>mykaul
marked this pull request as draft
January 29, 2026 16:23
mykaul
commented
Jan 29, 2026
Author
004080d caused a performance regression. I need to fix it (hopefully this weekend) |
8 tasks
mykaul
commented
Jan 29, 2026
Author
Replaced with #659 |
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
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Add slots definition to multiple classes in the project.
It started with one set (frames), then I did additional more, each separately, in theory, independently (but they were tested only one on top of the other).
Example from frame:
Pre-review checklist
./docs/source/.Fixes:annotations to PR description.