Hi Jeyhun, Thank you for driving this, it would be very useful optimization!
Sorry for joining the discussion now(I originally planned to reply earlier, but happened to be during my vacation). I have two questions: 1. Is this flip targeted only at batch scenarios or does it include streaming? (The flip and the discussion did not explicitly mention this, but in the draft pr, I only saw the implementation for batch scenarios https://github.com/apache/flink/pull/24437/files#diff-a6d71dd7d9bf0e7776404f54473b504e1de1240e93f820214fa5d1f082fb30c8 <https://github.com/apache/flink/pull/24437/files#diff-a6d71dd7d9bf0e7776404f54473b504e1de1240e93f820214fa5d1f082fb30c8%EF%BC%89> ) If we expect this also apply to streaming, then we need to consider the stricter shuffle restrictions of streaming compared to batch (if support is considered, more discussion is needed here, let’s not expand for now). If it only applies to batch, it is recommended to clarify in the flip. 2. In the current implementation, the optimized plan seems to have some problems. As described in the class comments: https://github.com/apache/flink/blob/d6e3b51fdb9a2e565709e8d7bc619234b3768ed1/flink-table/flink-table-planner/src/main/java/org/apache/flink/table/planner/plan/rules/physical/batch/RemoveRedundantShuffleRule.java#L60 BatchPhysicalHashAggregate (local) +- BatchPhysicalLocalHashAggregate (local) +- BatchPhysicalTableSourceScan The `BatchPhysicalLocalHashAggregate` here is redundant (in the case of one-phase hashAgg, localAgg is not necessary, which is the scenario currently handled by `RemoveRedundantLocalHashAggRule` and other rules). Also, in the draft pr, the optimization of `testShouldEliminatePartitioning1` & `testShouldEliminatePartitioning2` seems didn't take effect? https://github.com/apache/flink/blob/d6e3b51fdb9a2e565709e8d7bc619234b3768ed1/flink-table/flink-table-planner/src/test/resources/org/apache/flink/connector/file/table/BatchFileSystemTableSourceTest.xml#L38 In conjunction with question 2, I am wondering if we have a better choice (of course, not simply adding the current `PHYSICAL_OPT_RULES`'s `RemoveRedundantLocalXXRule`s to the `PHYSICAL_REWRITE`). For example, let the source actively provide some traits (including `FlinkRelDistribution` and `RelCollation`) to the planner. The advantage of doing this is to directly reuse the current shuffle remove optimization (as `FlinkExpandConversionRule` implemented), and according to the data distribution characteristics provided by the source, the planner may choose a physical operator with a cheaper costs (for example, according to `RelCollation`, the planner can use sortAgg, no need for a separate local sort operation). WDYT? Best, Lincoln Lee Jeyhun Karimov <je.kari...@gmail.com> 于2024年4月1日周一 18:00写道： > Hi everyone, > > Thanks for your valuable feedback! > > The discussion on this FLIP has been going on for a while. > I would like to start a vote after 48 hours. > > Please let me know if you have any concerns or any further > questions/comments. > > Regards, > Jeyhun > > > On Thu, Mar 21, 2024 at 6:01 PM Jeyhun Karimov <je.kari...@gmail.com> > wrote: > > > Hi Lorenzo, > > > > Thanks a lot for your comments. Please find my answers below: > > > > > > For the interface `SupportsPartitioning`, why returning `Optional`? > >> If one decides to implement that, partitions must exist (at maximum, > >> return and empty list). Returning `Optional` seem just to complicate the > >> logic of the code using that interface. > > > > > > - The reasoning behind the use of Optional is that sometimes (e.g., in > > HiveTableSource) the partitioning info is in catalog. > > Therefore, we return Optional.empty(), so that the list of partitions > is > > queried from the catalog. > > > > > > I foresee the using code doing something like: "if the source supports > >> partitioning, get the partitions, but if they don't exist, raise a > runtime > >> exception". Let's simply make that safe at compile time and guarantee > the > >> code that partitions exist. > > > > > > - Yes, once partitions cannot be found, neither from catalog nor from the > > interface implementation, then we raise an exception during query compile > > time. > > > > > > Another thing is that you show Hive-like partitioning in your FS > >> structure, do you think it makes sense to add a note about > auto-discovery > >> of partitions? > > > > > > - Yes, the FLIP contains just an example partitioning for filesystem > > connector. Each connector already "knows" about autodiscovery of its > > partitions. And we rely on this fact. > > For example, partition discovery is different between kafka and > > filesystem sources. So, we do not handle the manual discovery of > > partitions. Please correct me if I misunderstood your question. > > > > > > In other terms, it looks a bit counterintuitive that the user > implementing > >> the source has to specify which partitions exist statically (and they > can > >> change at runtime), while the source itself knows the data provider and > can > >> directly implement a method `discoverPartitions`. Then Flink would take > >> care of invoking that method when needed. > > > > > > We utilize table option SOURCE_MONITOR_INTERVAL to check whether > > partitions are static or not. So, a user still should give Flink a hint > > about partitions being static or not. With static partitions Flink can do > > more optimizations. > > > > Please let me know if my replies answer your questions or if you have > more > > comments. > > > > Regards, > > Jeyhun > > > > > > > > On Thu, Mar 21, 2024 at 10:03 AM <lorenzo.affe...@ververica.com> wrote: > > > >> Hello Jeyhun, > >> I really like the proposal and definitely makes sense to me. > >> > >> I have a couple of nits here and there: > >> > >> For the interface `SupportsPartitioning`, why returning `Optional`? > >> If one decides to implement that, partitions must exist (at maximum, > >> return and empty list). Returning `Optional` seem just to complicate the > >> logic of the code using that interface. > >> > >> I foresee the using code doing something like: "if the source supports > >> partitioning, get the partitions, but if they don't exist, raise a > runtime > >> exception". Let's simply make that safe at compile time and guarantee > the > >> code that partitions exist. > >> > >> Another thing is that you show Hive-like partitioning in your FS > >> structure, do you think it makes sense to add a note about > auto-discovery > >> of partitions? > >> > >> In other terms, it looks a bit counterintuitive that the user > >> implementing the source has to specify which partitions exist statically > >> (and they can change at runtime), while the source itself knows the data > >> provider and can directly implement a method `discoverPartitions`. Then > >> Flink would take care of invoking that method when needed. > >> On Mar 15, 2024 at 22:09 +0100, Jeyhun Karimov <je.kari...@gmail.com>, > >> wrote: > >> > >> Hi Benchao, > >> > >> Thanks for your comments. > >> > >> 1. What the parallelism would you take? E.g., 128 + 256 => 128? What > >> > >> if we cannot have a good greatest common divisor, like 127 + 128, > >> could we just utilize one side's pre-partitioned attribute, and let > >> another side just do the shuffle? > >> > >> > >> > >> There are two cases we need to consider: > >> > >> 1. Static Partition (no partitions are added during the query execution) > >> is > >> enabled AND both sources implement "SupportsPartitionPushdown" > >> > >> In this case, we are sure that no new partitions will be added at > runtime. > >> So, we have a chance equalize both sources' partitions and parallelism, > >> IFF > >> both sources implement "SupportsPartitionPushdown" interface. > >> To achieve so, first we will fetch the existing partitions from source1 > >> (say p_s1) and from source2 (say p_s2). > >> Then, we find the intersection of these two partition sets (say > >> p_intersect) and pushdown these partitions: > >> > >> SupportsPartitionPushDown::applyPartitions(p_intersect) // make sure > that > >> only specific partitions are read > >> SupportsPartitioning::applyPartitionedRead(p_intersect) // partitioned > >> read > >> with filtered partitions > >> > >> Lastly, we need to change the parallelism of 1) source1, 2) source2, and > >> 3) > >> all of their downstream operators until (and including) their first > common > >> ancestor (e.g., join) to be equal to the number of partitions (size of > >> p_intersect). > >> > >> 2. All other cases > >> > >> In all other cases, the parallelism of both sources and their downstream > >> operators until their common ancestor would be equal to the MIN(p_s1, > >> p_s2). > >> That is, minimum of the partition size of source1 and partition size of > >> source2 will be selected as the parallelism. > >> Coming back to your example, if source1 parallelism is 127 and source2 > >> parallelism is 128, then we will first check the partition size of > source1 > >> and source2. > >> Say partition size of source1 is 100 and partition size of source2 is > 90. > >> Then, we would set the parallelism for source1, source2, and all of > their > >> downstream operators until (and including) the join operator > >> to 90 (min(100, 90)). > >> We also plan to implement a cost based decision instead of the > rule-based > >> one (the ones explained above - MIN rule). > >> One possible result of the cost based estimation is to keep the > partitions > >> on one side and perform the shuffling on another source. > >> > >> > >> > >> 2. In our current shuffle remove design (FlinkExpandConversionRule), > >> > >> we don't consider parallelism, we just remove unnecessary shuffles > >> according to the distribution columns. After this FLIP, the > >> parallelism may be bundled with source's partitions, then how will > >> this optimization accommodate with FlinkExpandConversionRule, will you > >> also change downstream operator's parallelisms if we want to also > >> remove subsequent shuffles? > >> > >> > >> > >> > >> - From my understanding of FlinkExpandConversionRule, its removal logic > is > >> agnostic to operator parallelism. > >> So, if FlinkExpandConversionRule decides to remove a shuffle operation, > >> then this FLIP will search another possible shuffle (the one closest to > >> the > >> source) to remove. > >> If there is such an opportunity, this FLIP will remove the shuffle. So, > >> from my understanding FlinkExpandConversionRule and this optimization > rule > >> can work together safely. > >> Please correct me if I misunderstood your question. > >> > >> > >> > >> Regarding the new optimization rule, have you also considered to allow > >> > >> some non-strict mode like FlinkRelDistribution#requireStrict? For > >> example, source is pre-partitioned by a, b columns, if we are > >> consuming this source, and do a aggregate on a, b, c, can we utilize > >> this optimization? > >> > >> > >> > >> - Good point. Yes, there are some cases that non-strict mode will apply. > >> For example: > >> > >> - pre-partitioned columns and aggregate columns are the same but have > >> different order (e.g., source pre-partitioned w.r.t. a,b and aggregate > has > >> a GROUP BY b,a) > >> - columns in the Exchange operator is a list-prefix of pre-partitoned > >> columns of source (e.g., source is pre-partitioned w.r.t. a,b,c and > >> Exchange's partition columns are a,b) > >> > >> Please let me know if the above answers your questions or if you have > any > >> other comments. > >> > >> Regards, > >> Jeyhun > >> > >> On Thu, Mar 14, 2024 at 12:48 PM Benchao Li <libenc...@apache.org> > wrote: > >> > >> Thanks Jeyhun for bringing up this discussion, it is really exiting, > >> +1 for the general idea. > >> > >> We also introduced a similar concept in Flink Batch internally to cope > >> with bucketed tables in Hive, it is a very important improvement. > >> > >> > One thing to note is that for join queries, the parallelism of each > join > >> > source might be different. This might result in > >> > inconsistencies while using the pre-partitioned/pre-divided data > (e.g., > >> > different mappings of partitions to source operators). > >> > Therefore, it is the job of planner to detect this and adjust the > >> > parallelism. With that having in mind, > >> > the rest (how the split assigners perform) is consistent among many > >> > sources. > >> > >> > >> Could you elaborate a little more on this. I added my two cents here > >> about this part: > >> 1. What the parallelism would you take? E.g., 128 + 256 => 128? What > >> if we cannot have a good greatest common divisor, like 127 + 128, > >> could we just utilize one side's pre-partitioned attribute, and let > >> another side just do the shuffle? > >> 2. In our current shuffle remove design (FlinkExpandConversionRule), > >> we don't consider parallelism, we just remove unnecessary shuffles > >> according to the distribution columns. After this FLIP, the > >> parallelism may be bundled with source's partitions, then how will > >> this optimization accommodate with FlinkExpandConversionRule, will you > >> also change downstream operator's parallelisms if we want to also > >> remove subsequent shuffles? > >> > >> > >> Regarding the new optimization rule, have you also considered to allow > >> some non-strict mode like FlinkRelDistribution#requireStrict? For > >> example, source is pre-partitioned by a, b columns, if we are > >> consuming this source, and do a aggregate on a, b, c, can we utilize > >> this optimization? > >> > >> Jane Chan <qingyue....@gmail.com> 于2024年3月14日周四 15:24写道： > >> > >> > > >> > Hi Jeyhun, > >> > > >> > Thanks for your clarification. > >> > > >> > >> > > Once a new partition is detected, we add it to our existing mapping. > >> > >> Our > >> > >> > mapping looks like Map<Integer, Set<Integer>> > >> > >> subtaskToPartitionAssignment, > >> > >> > where it maps each source subtaskID to zero or more partitions. > >> > > >> > I understand your point. **It would be better if you could sync the > >> > content to the FLIP**. > >> > > >> > Another thing is I'm curious about what the physical plan looks like. > Is > >> > there any specific info that will be added to the table source (like > >> > filter/project pushdown)? It would be great if you could attach an > >> > >> example > >> > >> > to the FLIP. > >> > > >> > Bests, > >> > Jane > >> > > >> > On Wed, Mar 13, 2024 at 9:11 PM Jeyhun Karimov <je.kari...@gmail.com> > >> > >> wrote: > >> > >> > > >> > >> > > Hi Jane, > >> > > > >> > > Thanks for your comments. > >> > > > >> > > > >> > > 1. Concerning the `sourcePartitions()` method, the partition > >> > >> information > >> > >> > > > returned during the optimization phase may not be the same as the > >> > >> > > partition > >> > >> > > > information during runtime execution. For long-running jobs, > >> > >> partitions > >> > >> > > may > >> > >> > > > be continuously created. Is this FLIP equipped to handle > scenarios? > >> > >> > > > >> > > > >> > > - Good point. This scenario is definitely supported. > >> > > Once a new partition is added, or in general, new splits are > >> > > discovered, > >> > > PartitionAwareSplitAssigner::addSplits(Collection<FileSourceSplit> > >> > > newSplits) > >> > > method will be called. Inside that method, we are able to detect if > a > >> > >> split > >> > >> > > belongs to existing partitions or there is a new partition. > >> > > Once a new partition is detected, we add it to our existing mapping. > >> > >> Our > >> > >> > > mapping looks like Map<Integer, Set<Integer>> > >> > >> subtaskToPartitionAssignment, > >> > >> > > where > >> > > it maps each source subtaskID to zero or more partitions. > >> > > > >> > > 2. Regarding the `RemoveRedundantShuffleRule` optimization rule, I > >> > >> > > > understand that it is also necessary to verify whether the hash > key > >> > >> > > within > >> > >> > > > the Exchange node is consistent with the partition key defined in > >> the > >> > >> > > table > >> > >> > > > source that implements `SupportsPartitioning`. > >> > >> > > > >> > > > >> > > - Yes, I overlooked that point, fixed. Actually, the rule is much > >> > > complicated. I tried to simplify it in the FLIP. Good point. > >> > > > >> > > > >> > > 3. Could you elaborate on the desired physical plan and integration > >> > >> with > >> > >> > > > `CompiledPlan` to enhance the overall functionality? > >> > >> > > > >> > > > >> > > - For compiled plan, PartitioningSpec will be used, with a json tag > >> > > "Partitioning". As a result, in the compiled plan, the source > operator > >> > >> will > >> > >> > > have > >> > > "abilities" : [ { "type" : "Partitioning" } ] as part of the > compiled > >> > >> plan. > >> > >> > > More about the implementation details below: > >> > > > >> > > -------------------------------- > >> > > PartitioningSpec class > >> > > -------------------------------- > >> > > @JsonTypeName("Partitioning") > >> > > public final class PartitioningSpec extends SourceAbilitySpecBase { > >> > > // some code here > >> > > @Override > >> > > public void apply(DynamicTableSource tableSource, > >> > >> SourceAbilityContext > >> > >> > > context) { > >> > > if (tableSource instanceof SupportsPartitioning) { > >> > > ((SupportsPartitioning<?>) > >> > >> tableSource).applyPartitionedRead(); > >> > >> > > } else { > >> > > throw new TableException( > >> > > String.format( > >> > > "%s does not support > >> > >> SupportsPartitioning.", > >> > >> > > tableSource.getClass().getName())); > >> > > } > >> > > } > >> > > // some code here > >> > > } > >> > > > >> > > -------------------------------- > >> > > SourceAbilitySpec class > >> > > -------------------------------- > >> > > @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = > >> > > JsonTypeInfo.As.PROPERTY, property = "type") > >> > > @JsonSubTypes({ > >> > > @JsonSubTypes.Type(value = FilterPushDownSpec.class), > >> > > @JsonSubTypes.Type(value = LimitPushDownSpec.class), > >> > > @JsonSubTypes.Type(value = PartitionPushDownSpec.class), > >> > > @JsonSubTypes.Type(value = ProjectPushDownSpec.class), > >> > > @JsonSubTypes.Type(value = ReadingMetadataSpec.class), > >> > > @JsonSubTypes.Type(value = WatermarkPushDownSpec.class), > >> > > @JsonSubTypes.Type(value = SourceWatermarkSpec.class), > >> > > @JsonSubTypes.Type(value = AggregatePushDownSpec.class), > >> > > + @JsonSubTypes.Type(value = PartitioningSpec.class) > >> > >> // > >> > >> > > new added > >> > > > >> > > > >> > > > >> > > Please let me know if that answers your questions or if you have > other > >> > > comments. > >> > > > >> > > Regards, > >> > > Jeyhun > >> > > > >> > > > >> > > On Tue, Mar 12, 2024 at 8:56 AM Jane Chan <qingyue....@gmail.com> > >> > >> wrote: > >> > >> > > > >> > >> > > > Hi Jeyhun, > >> > > > > >> > > > Thank you for leading the discussion. I'm generally +1 with this > >> > >> > > proposal, > >> > >> > > > along with some questions. Please see my comments below. > >> > > > > >> > > > 1. Concerning the `sourcePartitions()` method, the partition > >> > >> information > >> > >> > > > returned during the optimization phase may not be the same as the > >> > >> > > partition > >> > >> > > > information during runtime execution. For long-running jobs, > >> > >> partitions > >> > >> > > may > >> > >> > > > be continuously created. Is this FLIP equipped to handle > scenarios? > >> > > > > >> > > > 2. Regarding the `RemoveRedundantShuffleRule` optimization rule, I > >> > > > understand that it is also necessary to verify whether the hash > key > >> > >> > > within > >> > >> > > > the Exchange node is consistent with the partition key defined in > >> the > >> > >> > > table > >> > >> > > > source that implements `SupportsPartitioning`. > >> > > > > >> > > > 3. Could you elaborate on the desired physical plan and > integration > >> > >> with > >> > >> > > > `CompiledPlan` to enhance the overall functionality? > >> > > > > >> > > > Best, > >> > > > Jane > >> > > > > >> > > > On Tue, Mar 12, 2024 at 11:11 AM Jim Hughes > >> > >> <jhug...@confluent.io.invalid > >> > >> > > > > >> > > > wrote: > >> > > > > >> > >> > > > > Hi Jeyhun, > >> > > > > > >> > > > > I like the idea! Given FLIP-376[1], I wonder if it'd make sense > to > >> > > > > generalize FLIP-434 to be about "pre-divided" data to cover > >> > >> "buckets" > >> > >> > > and > >> > >> > > > > "partitions" (and maybe even situations where a data source is > >> > >> > > > partitioned > >> > >> > > > > and bucketed). > >> > > > > > >> > > > > Separate from that, the page mentions TPC-H Q1 as an example. > For > >> > >> a > >> > >> > > > join, > >> > >> > > > > any two tables joined on the same bucket key should provide a > >> > >> concrete > >> > >> > > > > example of a join. Systems like Kafka Streams/ksqlDB call this > >> > > > > "co-partitioning"; for those systems, it is a requirement placed > >> > >> on the > >> > >> > > > > input sources. For Flink, with FLIP-434, the proposed planner > rule > >> > > > > could remove the shuffle. > >> > > > > > >> > > > > Definitely a fun idea; I look forward to hearing more! > >> > > > > > >> > > > > Cheers, > >> > > > > > >> > > > > Jim > >> > > > > > >> > > > > > >> > > > > 1. > >> > > > > > >> > > > > > >> > >> > > > > >> > >> > > > >> > >> > >> > https://cwiki.apache.org/confluence/display/FLINK/FLIP-376%3A+Add+DISTRIBUTED+BY+clause > >> > >> > > > > 2. > >> > > > > > >> > > > > > >> > >> > > > > >> > >> > > > >> > >> > >> > https://docs.ksqldb.io/en/latest/developer-guide/joins/partition-data/#co-partitioning-requirements > >> > >> > > > > > >> > > > > On Sun, Mar 10, 2024 at 3:38 PM Jeyhun Karimov < > >> > >> je.kari...@gmail.com> > >> > >> > > > > wrote: > >> > > > > > >> > >> > > > > > Hi devs, > >> > > > > > > >> > > > > > I’d like to start a discussion on FLIP-434: Support > >> > >> optimizations for > >> > >> > > > > > pre-partitioned data sources [1]. > >> > > > > > > >> > > > > > The FLIP introduces taking advantage of pre-partitioned data > >> > >> sources > >> > >> > > > for > >> > >> > > > > > SQL/Table API (it is already supported as experimental feature > >> in > >> > > > > > DataStream API [2]). > >> > > > > > > >> > > > > > > >> > > > > > Please find more details in the FLIP wiki document [1]. > >> > > > > > Looking forward to your feedback. > >> > > > > > > >> > > > > > Regards, > >> > > > > > Jeyhun > >> > > > > > > >> > > > > > [1] > >> > > > > > > >> > > > > > > >> > >> > > > > > >> > >> > > > > >> > >> > > > >> > >> > >> > https://cwiki.apache.org/confluence/display/FLINK/FLIP-434%3A+Support+optimizations+for+pre-partitioned+data+sources > >> > >> > > > > > [2] > >> > > > > > > >> > > > > > > >> > >> > > > > > >> > >> > > > > >> > >> > > > >> > >> > >> > https://nightlies.apache.org/flink/flink-docs-master/docs/dev/datastream/experimental/ > >> > >> > > > > > > >> > >> > > > > > >> > >> > > > > >> > >> > > > >> > >> > >> > >> > >> -- > >> > >> Best, > >> Benchao Li > >> > >> >
