Thanks Jark, fair enough, I will update the FLIP accordingly. Best regards, Jing
On Fri, Jul 22, 2022 at 6:07 AM Jark Wu <imj...@gmail.com> wrote: > Hi Jing, > > I have some concerns about the isBulkGetSupported() approach. > 1. Catalog developers need to learn the contract between > `isBulkGetSupported()` and bulk get methods > 2. The contract of isBulkGetSupported() is fragile. Because developers may > forget to override > `isBulkGetSupported` and be confused the implemented bulk methods are > not called. > 3. It's not extensible when we have more bulk methods in future releases. > > I have discussed with Godfrey offline and come up with (maybe) a better > solution. > The default implementation of bulk methods should iterate on the single get > method and wrap the result. > In this way, the planner can always call the bulk methods no matter whether > the Catalog implements the bulk method or not. > This can also address the above concerns. > > What do you think? > > Best, > Jark > > > > > > On Thu, 21 Jul 2022 at 21:13, Jing Ge <j...@ververica.com> wrote: > > > Thanks Jingsong and Jark. I will create another FLIP to cover the > > optimization topic that partitions and partition stats could be fetched > > from catalog in one single call. > > > > Thanks for the hint w.r.t. the compatibility issue. I have updated the > FLIP > > to provide all methods as default interface methods. A new > > method isBulkGetSupported() has been added. Would you like to take a look > > at it again? > > > > Best regards, > > Jing > > > > On Wed, Jul 20, 2022 at 11:23 AM Jark Wu <imj...@gmail.com> wrote: > > > > > I agree with Jingsong. > > > > > > There are use cases to get partitions and partition stats in a single > > call > > > to reduce the IO cost. > > > For example, extending Catalog#listPartitions to > > > Catalog#listPartitionsWithStats, > > > and extending Catalog#listPartitionsByFilter to > > > Catalog#listPartitionsWithStatsByFilter. > > > This allows the planner to choose the cheapest one to request > partitions > > > and stats. > > > But of course, this can be introduced in the future. > > > > > > Besides, I found that the FLIP doesn't cover compatibility properly. > > > Introducing two new methods in Catalog > > > will break all third-party catalog implementations. On the other hand, > > the > > > planner should > > > have a way to identify whether the Catalog supports bulk get. > > > > > > Best, > > > Jark > > > > > > On Wed, 20 Jul 2022 at 16:43, Jingsong Li <jingsongl...@gmail.com> > > wrote: > > > > > > > Thanks for your reply. > > > > > > > > - Consider bulkGetPartitionStatistics, partition statistics are > > > > already in HiveMetastoreClient.listPartitions. But on our side, we > > > > need Catalog.getPartitions first, and then > > > > Catalog.bulkGetPartitionStatistics. > > > > > > > > - Consider bulkGetPartitionColumnStatistics, yes, as you said, we > need > > > it. > > > > > > > > But to unify them, the current FLIP is acceptable. > > > > > > > > Best, > > > > Jingsong > > > > > > > > On Wed, Jul 20, 2022 at 3:57 PM Jing Ge <j...@ververica.com> wrote: > > > > > > > > > > Hi Jingsong, > > > > > > > > > > Thanks for clarifying it. Are you suggesting a new method or > changing > > > the > > > > > name of the methods described in the FLIP? > > > > > Please see my answers and further questions below. > > > > > > > > > > Best regards, > > > > > Jing > > > > > > > > > > On Wed, Jul 20, 2022 at 4:28 AM Jingsong Li < > jingsongl...@gmail.com> > > > > wrote: > > > > > > > > > > > Hi Jing, > > > > > > > > > > > > I understand that the statistics for partitions are currently > only > > > > > > used by Hive, so we can look at the Hive implementation: > > > > > > > > > > > > See HiveCatalog.getPartitionStatistics. > > > > > > To get the statistics, we actually get them from the > > > > > > org.apache.hadoop.hive.metastore.api.Partition object. > > > > > > > > > > > > > > > > Correct. Both new methods will do the same thing but in bulk mode. > > > > > > > > > > > > > > > > > > > > > > According to HiveMetastore's API, partition-related operations > > > > > > actually get the partition as well as the statistics information. > > > > > > > > > > > > > > > > I am not really sure which API it is. Methods in > HiveMetaStoreClient > > > that > > > > > are dealing with partitions will either return partitions or > > > > statisticObjs, > > > > > e.g. listPartitions(...) or getPartitionColumnStatistics(...) > > > > > > > > > > > > > > > > > > > > > > So if the current partition statistics are just for Hive, can we > > > > > > consider unifying it with Hive? > > > > > > > > > > > > > > > > Yes, we are on the same page. This FLIP is trying to unify it with > > > > > HiveMetaStoreClient. > > > > > > > > > > > > > > > > > > > > > > For example, in PushPartitionIntoTableSourceScanRule, just use > > > > > > `listPartitionWithStats`, and adjust table statistics from > > > partitions. > > > > > > > > > > > > > > > > Does the bulkGetPartitionStatistics work in this case too? Or, do > you > > > > mean > > > > > you need both partitions and related statistics returned by a new > > > method > > > > > called `listPartitionWithStats`? > > > > > > > > > > > > > > > > Best, > > > > > > Jingsong > > > > > > > > > > > > On Tue, Jul 19, 2022 at 8:44 PM Jing Ge <j...@ververica.com> > > wrote: > > > > > > > > > > > > > > Thanks Jingsong for the suggestion. > > > > > > > > > > > > > > Do you mean using a different naming convention? There is a > > thought > > > > and > > > > > > > description in the FLIP about using "list" or "bulkGet": > > > > > > > > > > > > > > - bulkGetPartitionStatistics(...) has been chosen over > > > > > > > listPartitionStatistics(...), because, comparing to database > > and > > > > > > partition > > > > > > > that are static and can be listed, statistics are more > dynamic > > > and > > > > > > will > > > > > > > need more computation logic to create, therefore using "get" > > is > > > > > > > semantically more feasible than list. The "bulk" gives users > > the > > > > hint > > > > > > that > > > > > > > this method will work in the bulk mode and return a > collection > > > of > > > > > > instances. > > > > > > > > > > > > > > > > > > > > > As a reference, we can see that no method in MetaStoreClient, > > that > > > > > > > calculates statistics, uses the "list" naming convention. > > > > > > > > > > > > > > Best regards, > > > > > > > Jing > > > > > > > > > > > > > > On Fri, Jul 15, 2022 at 5:38 AM Jingsong Li < > > > jingsongl...@gmail.com> > > > > > > wrote: > > > > > > > > > > > > > > > Thanks for starting this discussion. > > > > > > > > > > > > > > > > Have we considered introducing a listPartitionWithStats() in > > > > Catalog? > > > > > > > > > > > > > > > > Best, > > > > > > > > Jingsong > > > > > > > > > > > > > > > > On Fri, Jul 15, 2022 at 10:08 AM Jark Wu <imj...@gmail.com> > > > wrote: > > > > > > > > > > > > > > > > > > Hi Jing, > > > > > > > > > > > > > > > > > > Thanks for starting this discussion. The bulk fetch is a > > great > > > > > > > > improvement > > > > > > > > > for the optimizer. > > > > > > > > > The FLIP looks good to me. > > > > > > > > > > > > > > > > > > Best, > > > > > > > > > Jark > > > > > > > > > > > > > > > > > > On Fri, 8 Jul 2022 at 17:36, Jing Ge <j...@ververica.com> > > > wrote: > > > > > > > > > > > > > > > > > > > Hi devs, > > > > > > > > > > > > > > > > > > > > After having multiple discussions with Jark and Goldfrey, > > I'd > > > > like > > > > > > to > > > > > > > > start > > > > > > > > > > a discussion on the mailing list w.r.t. FLIP-247[1], > which > > > will > > > > > > > > > > significantly improve the performance by providing the > bulk > > > > fetch > > > > > > > > > > capability for table and column statistics. > > > > > > > > > > > > > > > > > > > > Currently the statistics information about tables can > only > > be > > > > > > fetched > > > > > > > > from > > > > > > > > > > the catalog by each given partition iteratively. Since > > > getting > > > > > > > > statistics > > > > > > > > > > information from catalogs is a very heavy operation, in > > order > > > > to > > > > > > > > improve > > > > > > > > > > the query performance, we’d better provide functionality > to > > > > fetch > > > > > > the > > > > > > > > > > statistics information of a table for all given > partitions > > in > > > > one > > > > > > shot. > > > > > > > > > > > > > > > > > > > > Based on the manual performance test, for 2000 > partitions, > > > the > > > > cost > > > > > > > > will be > > > > > > > > > > improved from 10s to 2s. The improvement result is 500%. > > > > > > > > > > > > > > > > > > > > [1] > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-247%3A+Bulk+fetch+of+table+and+column+statistics+for+given+partitions > > > > > > > > > > > > > > > > > > > > Best regards, > > > > > > > > > > Jing > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
