stream2000 commented on code in PR #8062: URL: https://github.com/apache/hudi/pull/8062#discussion_r1211233703
########## rfc/rfc-65/rfc-65.md: ########## @@ -0,0 +1,110 @@ +## Proposers +- @stream2000 +- @hujincalrin +- @huberylee +- @YuweiXiao +## Approvers +## Status +JIRA: [HUDI-5823](https://issues.apache.org/jira/browse/HUDI-5823) +## Abstract +In some classic hudi use cases, users partition hudi data by time and are only interested in data from a recent period of time. The outdated data is useless and costly, we need a TTL(Time-To-Live) management mechanism to prevent the dataset from growing infinitely. +This proposal introduces Partition TTL Management policies to hudi, people can config the policies by table config directly or by call commands. With proper configs set, Hudi can find out which partitions are outdated and delete them. +## Background +TTL management mechanism is an important feature for databases. Hudi already provides a delete_partition interface to delete outdated partitions. However, users still need to detect which partitions are outdated and call `delete_partition` manually, which means that users need to define and implement some kind of TTL policies and maintain proper statistics to find expired partitions by themself. As the scale of installations grew, it's more important to implement a user-friendly TTL management mechanism for hudi. +## Implementation +There are 3 components to implement Partition TTL Management + +- TTL policy definition & storage +- Partition statistics for TTL management +- Appling policies +### TTL Policy Definition +We have three main considerations when designing TTL policy: + +1. User hopes to manage partition TTL not only by expired time but also by sub-partitions count and sub-partitions size. So we need to support the following three different TTL policy types. + 1. **KEEP_BY_TIME**. Partitions will expire N days after their last modified time. + 2. **KEEP_BY_COUNT**. Keep N sub-partitions for a high-level partition. When sub partition count exceeds, delete the partitions with smaller partition values until the sub-partition count meets the policy configuration. + 3. **KEEP_BY_SIZE**. Similar to KEEP_BY_COUNT, but to ensure that the sum of the data size of all sub-partitions does not exceed the policy configuration. +2. User need to set different policies for different partitions. For example, the hudi table is partitioned by two fields (user_id, ts). For partition(user_id='1'), we set the policy to keep 100G data for all sub-partitions, and for partition(user_id='2') we set the policy that all sub-partitions will expire 10 days after their last modified time. +3. It's possible that there are a lot of high-level partitions in the user's table, and they don't want to set TTL policies for all the high-level partitions. So we need to provide a default policy mechanism so that users can set a default policy for all high-level partitions and add some explicit policies for some of them if needed. Explicit policies will override the default policy. + +So here we have the TTL policy definition: +```java +public class HoodiePartitionTTLPolicy { + public enum TTLPolicy { + KEEP_BY_TIME, KEEP_BY_SIZE, KEEP_BY_COUNT + } + + // Partition spec for which the policy takes effect + private String partitionSpec; + + private TTLPolicy policy; + + private long policyValue; +} +``` + +### User Interface for TTL policy +Users can config partition TTL management policies through SparkSQL Call Command and through table config directly. Assume that the user has a hudi table partitioned by two fields(user_id, ts), he can config partition TTL policies as follows. + +```sql +-- Set default policy for all user_id, which keeps the data for 30 days. +call add_ttl_policy(table => 'test', partitionSpec => 'user_id=*/', policy => 'KEEP_BY_TIME', policyValue => '30'); + +--For partition user_id=1/, keep 10 sub partitions. +call add_ttl_policy(table => 'test', partitionSpec => 'user_id=1/', policy => 'KEEP_BY_COUNT', policyValue => '10'); + +--For partition user_id=2/, keep 100GB data in total +call add_ttl_policy(table => 'test', partitionSpec => 'user_id=2/', policy => 'KEEP_BY_SIZE', policyValue => '107374182400'); + +--For partition user_id=3/, keep the data for 7 day. +call add_ttl_policy(table => 'test', partitionSpec => 'user_id=3/', policy => 'KEEP_BY_TIME', policyValue => '7'); + +-- Show all the TTL policies including default and explicit policies +call show_ttl_policies(table => 'test'); +user_id=*/ KEEP_BY_TIME 30 +user_id=1/ KEEP_BY_COUNT 10 +user_id=2/ KEEP_BY_SIZE 107374182400 +user_id=3/ KEEP_BY_TIME 7 +``` + +### Storage for TTL policy +The partition TTL policies will be stored in `hoodie.properties`since it is part of table metadata. The policy configs in `hoodie.properties`are defined as follows. Explicit policies are defined using a JSON array while default policy is defined using separate configs. + +```properties +# Default TTL policy definition +hoodie.partition.ttl.management.default.policy=KEEP_BY_TIME +hoodie.partition.ttl.management.default.fields=user_id +hoodie.partition.ttl.management.default.policy.value=30 + +# Explicit TTL policy definition +hoodie.partition.ttl.policies=[{"partitionSpec"\:"user_id\=2/","policy"\:"KEEP_BY_SIZE","policyValue"\:107374182400},{"partitionSpec"\:"user_id\=1/","policy"\:"KEEP_BY_COUNT","policyValue"\:10},{"partitionSpec"\:"user_id\=3/","policy"\:"KEEP_BY_TIME","policyValue"\:7}] +``` + +### Partition Statistics +#### Partition Statistics Entity +We need need to maintain a partition stat for every partition in the hudi table. The stat will contain three fields: + +- RelativePath. Relative path to the base path, or we can call it PartitionPath. +- LastModifiedTime. Last instant time that modified the partition. We need this information to support the `KEEP_BY_TIME` policy. +- PartitionTotalFileSize. The sum of all valid data file sizes in the partition, to support the `KEEP_BY_SIZE`policy. We calculate only the latest file slices in the partition currently. +#### Gathering Partition Statistics +The simplest way to get partition statistics is that reading information from the metadata table. However, the write and compaction of the metadata table will severely affect the throughput and latency of data ingestion. So we design an asynchronous partition statistics as follows. + +- At the first time we gather the partitions statistics, we list all partitions in the hudi table and calculate `PartitionTotalFileSize`and `LastModifiedTime`for each partition. Then we store all the partition stats in persistent storage along with the instant time we do the stats. For example, we can directly store all partition stats in a JSON file whose file name contains the instant time. Review Comment: > Why a separate json file? Why not augment the delete partition commit metadata to have all this info? In our case, the number of partitions is so large that we can only get partition stats in an incremental manner. So we introduce a service that will periodically gather stats of modified partitions from timeline metadata and store the stats in the filesystem to accelerate ttl management. However, when the partition number is not so large, listing all partitions to get partition stats is acceptable. So the JSON file can be deemed as a kind of optimization and we may not need it in the first community version. -- 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]
