- 08 1月, 2021 10 次提交
-
-
由 Yuan Mei 提交于
-
由 V1ncentzzZ 提交于
This closes #14508
-
由 godfreyhe 提交于
[FLINK-20856][table-planner-blink] Introduce StreamPhysicalPythonGroupWindowAggregate, and make StreamExecPythonGroupWindowAggregate only extended from ExecNode This closes #14569
-
由 godfreyhe 提交于
[FLINK-20856][table-planner-blink] Introduce StreamPhysicalGroupWindowAggregate & StreamPhysicalGroupWindowTableAggregate, and make StreamExecGroupWindowAggregate only extended from ExecNode This closes #14569
-
由 Yangze Guo 提交于
[FLINK-20836][runtime] Pass total resource and default slot resource in registering TaskManager to SlotManager This closes #14561
-
由 Yangze Guo 提交于
[FLINK-20836][runtime] Add total resource profile and default slot resource profile to WorkerRegistration
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
FLINK-20717 introduced a bug, where any time spent on processing mails, when task is idle or back pressured, will be accounted to idle or back pressured time instead of the busy time. The fix is to assign idle or back pressure timer to the suspenssion marker and pause this timer when MailboxProcessor is running mails.
-
由 Timo Walther 提交于
All tests in modules apart from the Blink planner/runtime module have been updated. Otherwise we use a JUnit rule to make the migration of the remaining tests incremental. This closes #14568.
-
由 Timo Walther 提交于
This updates the Row.toString method to provide a good summary string. In particular it fixes the following issues: Changeflag: According to FLIP-95, a row describes an entry in a changelog. Therefore, it should visible whether a row is an insert, delete, or update change. Now indicated with +I, -D, +U, -U. Nested rows: In the old implementation it was not visible whether nested rows exist or not due to missing start/end boundaries. Now indicated with [...] or {...}. Positioned rows vs. named rows: According to FLIP-136, it should be visible whether a row operates in name-based or position-based field mode. Now indicated with [...] or {...}. Nested arrays in maps and lists: In the old implementation arrays in maps or lists could not be represented. Wrong formatting: Most programming languages use a space after a comma. This is an incompatible change. If the legacy representation is still required for tests, the old behavior can be restored via the flag RowUtils.USE_LEGACY_TO_STRING for the local JVM. However, relying on the row's string representation for tests is not a good idea in general as field data types are not verified.
-
- 07 1月, 2021 22 次提交
-
-
由 Jiangjie (Becket) Qin 提交于
-
由 godfreyhe 提交于
[FLINK-20738][table-planner-blink] Introduce BatchPhysicalPythonGroupAggregate, and make BatchExecPythonGroupAggregate only extended from ExecNode This closes #14562
-
由 godfreyhe 提交于
[FLINK-20738][table-planner-blink] Introduce BatchPhysicalLocalSortAggregate, and make BatchPhysicalLocalSortAggregate only extended from FlinkPhysicalRel This closes #14562
-
由 godfreyhe 提交于
[FLINK-20738][table-planner-blink] Introduce BatchPhysicalSortAggregate, and make BatchExecSortAggregate only extended from ExecNode This closes #14562
-
由 godfreyhe 提交于
[FLINK-20738][table-planner-blink] Introduce BatchPhysicalLocalHashAggregate, and make BatchPhysicalLocalHashAggregate only extended from FlinkPhysicalRel This closes #14562
-
由 godfreyhe 提交于
[FLINK-20738][table-planner-blink] Introduce BatchPhysicalHashAggregate, and make BatchExecHashAggregate only extended from ExecNode This closes #14562
-
由 godfreyhe 提交于
[FLINK-20738][table-planner-blink] Rename BatchExecGroupAggregateBase to BatchPhysicalGroupAggregateBase and do some refactoring This closes #14562
-
由 Chesnay Schepler 提交于
-
由 Dawid Wysakowicz 提交于
-
由 Aljoscha Krettek 提交于
-
由 Aljoscha Krettek 提交于
This adds an ITCase because we need to check that all the components work together and that the wiring works correctly. This uses the previously added funtionality to specify that given inputs should be processed before other inputs and should not be sorted.
-
由 Aljoscha Krettek 提交于
This doesn't change the actual behavior, we still set the same "sorted" setting on both inputs. We will add tests and actually change the behavior in a follow-up commit.
-
由 Aljoscha Krettek 提交于
We allow the operation but it is a no-op because there never is any keyed state because we process all broadcast input before starting to process the keyed/other side.
-
由 Aljoscha Krettek 提交于
This will allow processing the broadcast side of a broadcast operator first, before processing the keyed side that requires sorting for stateful BATCH execution. For now, the wiring from the API is not there, this will be added in follow-up changes.
-
由 Aljoscha Krettek 提交于
Note: Broadcast operations in BATCH mode don't yet work with this change. This needs follow-up changes from later commits. We just lay the groundwork here and keep the same functionality. Before, we were creating operators in BroadcastConnectedStreams eagerly. Now, the transformation holds the user function and we add a Translator that creates the "physical" operators when translating the graph of Transformations. We do this so that we can translate differently based on whether we're in BATCH or STREAMING mode.
-
由 Aljoscha Krettek 提交于
Because that's what it does.
-
由 Dian Fu 提交于
-
由 龙三 提交于
[FLINK-20766][table-planner-blink] Introduce BatchPhysicalSortLimit, and make BatchExecSortLimit only extended from ExecNode This closes #14502
-
由 龙三 提交于
[FLINK-20766][table-planner-blink] Introduce BatchPhysicalSort, and make BatchExeSort only extended from ExecNode This closes #14502
-
由 龙三 提交于
[FLINK-20766][table-planner-blink] Introduce StreamPhysicalTemporalSort, and make StreamExecTemporalSort only extended from ExecNode. This closes #14502
-
由 龙三 提交于
[FLINK-20766][table-planner-blink] Introduce StreamPhysicalSortLimit, and make StreamExecSortLimit only extends from ExecNode. This closes #14502
-
由 龙三 提交于
[FLINK-20766][table-planner-blink] Introduce StreamPhysicalSort, and make StreamExecSort only extended from ExecNode This closes #14502
-
- 06 1月, 2021 8 次提交
-
-
由 龙三 提交于
[FLINK-20782][table-planner-blink] Introduce BatchPhysicalRank, and make BatchExecRank only extended from ExecNode This closes #14506
-
由 Piotr Nowojski 提交于
In particularly, if task is idling forever, as there are no new records incomming previous version would report idleTime as 0% and busyTime as 100% which is incorrect. In this version, idleTime metric is aware that idling period has started and can take that into account when updating it's value.
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
It's defined as inverted value of idleTimeMsPerSecond
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
-
由 Piotr Nowojski 提交于
-