danny0405 commented on code in PR #11559: URL: https://github.com/apache/hudi/pull/11559#discussion_r1699354428
########## rfc/rfc-80/rfc-80.md: ########## @@ -0,0 +1,187 @@ +<!-- + Licensed to the Apache Software Foundation (ASF) under one or more + contributor license agreements. See the NOTICE file distributed with + this work for additional information regarding copyright ownership. + The ASF licenses this file to You under the Apache License, Version 2.0 + (the "License"); you may not use this file except in compliance with + the License. You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. +--> +# RFC-80: Support column families for wide tables + +## Proposers +- @xiarixiaoyao +- @wombatu-kun + +## Approvers + - @vinothchandar + - @danny0405 + - @beyond1920 + +## Status + +JIRA: https://issues.apache.org/jira/browse/HUDI-7947 + +## Abstract + +In streaming processing, there are often scenarios where the table is widened. The current mainstream real-time wide table concatenation is completed through Flink's multi-layer join; +Flink's join will cache a large amount of data in the state backend. As the data set increases, the pressure on the Flink task state backend will gradually increase, and may even become unavailable. +In multi-layer join scenarios, this problem is more obvious. +1.x also supports partial updates being encoded in logfiles. That should be able to handle this scenario. But even with partial-update, the column families will reduce write amplification on compaction. + +So, main gains of clustering columns for wide tables are: +Write performance: +- Writing is similar to ordinary bucket writing, but it involves splitting column clusters and sorting. Therefore, the writing performance of full data is 10% lower than that of native bucket writing. +- However, if only some columns are updated among a large number of columns, the writing efficiency is much faster than that of non-column clustered tables. (Note: 1.0 introduces a partial update functionality, that can also avoid such writing costs, by writing only the changed columns) + +Read performance: +Since the data is already sorted when it is written, the SortMerge method can be used directly to merge the data; compared with the native bucket data reading performance is improved a lot, and the memory consumption is reduced significantly. + +Compaction performance: +The logic of compaction and reading is the same. Compaction costs across column families is where there real savings are. + +The log merge we can make it pluggable to decide between hash or sort merge - we need to introduce new log headers or standard mechanism for merging to determine if base file or log files are sorted. + +## Background +Currently, Hudi organizes data according to fileGroup granularity. The fileGroup is further divided into column clusters to introduce the columnFamily concept. +The organizational form of Hudi files is divided according to the following rules: +The data in the partition is divided into buckets according to hash (each bucket maps to a file group); the files in each bucket are divided according to columnFamily; multiple colFamily files in the bucket form a completed fileGroup; when there is only one columnFamily, it degenerates into the native Hudi bucket table. + + + +## Implementation +This feature should be implemented for both Spark and Flink. So, a table written by Flink this way, also can be read by Spark. + +### Constraints and Restrictions +1. The overall design relies on the non-blocking concurrent writing feature of Hudi 1.0. +2. Lower version Hudi cannot read and write column family tables. +3. Only MOR bucketed tables support setting column families. + MOR+Bucket is more suitable because it has higher write performance, but this does not mean that column family is incompatible with other indexes and cow tables. +4. Column families do not support repartitioning and renaming. +5. Schema evolution does not take effect on the current column family table. + Not supporting Schema evolution does not mean users can not add/delete columns in their table, they just need to do it explicitly. +6. Like native bucket tables, clustering operations are not supported. + +### Model change +After the column family is introduced, the storage structure of the entire Hudi bucket table changes: + + + +The bucket is divided into multiple columnFamilies by column cluster. When columnFamily is 1, it will automatically degenerate into the native bucket table. + + + +### Proposed Storage Format Changes +After splitting the fileGroup by columnFamily, the naming rules for base files and log files change. We add the cfName suffix at the end of all file names to facilitate Hudi itself to distinguish column families. If it's not present, we assume default column family. +So, new file name templates will be as follows: +- Base file: [file_id]\_[write_token]\_[begin_time]**[#cfName]**.[extension] Review Comment: I guess the `cfName` is separated with a `_` prefix right? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
