codope commented on a change in pull request #4640: URL: https://github.com/apache/hudi/pull/4640#discussion_r822835191
########## File path: rfc/rfc-45/rfc-45.md ########## @@ -0,0 +1,229 @@ +<!-- + 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-45: Asynchronous Metadata Indexing + +## Proposers + +- @codope +- @manojpec + +## Approvers + +- @nsivabalan +- @vinothchandar + +## Status + +JIRA: [HUDI-2488](https://issues.apache.org/jira/browse/HUDI-2488) + +## Abstract + +Metadata indexing (aka metadata bootstrapping) is the process of creation of one +or more metadata-based indexes, e.g. data partitions to files index, that is +stored in Hudi metadata table. Currently, the metadata table (referred as MDT +hereafter) supports single partition which is created synchronously with the +corresponding data table, i.e. commits are first applied to metadata table +followed by data table. Our goal for MDT is to support multiple partitions to +boost the performance of existing index and records lookup. However, the +synchronous manner of metadata indexing is not very scalable as we add more +partitions to the MDT because the regular writers (writing to the data table) +have to wait until the MDT commit completes. In this RFC, we propose a design to +support asynchronous metadata indexing. + +## Background + +We can read more about the MDT design +in [RFC-15](https://cwiki.apache.org/confluence/display/HUDI/RFC+-+15%3A+HUDI+File+Listing+Improvements) +. Here is a quick summary of the current state (Hudi v0.10.1). MDT is an +internal Merge-on-Read (MOR) table that has a single partition called `files` +which stores the data partitions to files index that is used in file listing. +MDT is co-located with the data table (inside `.hoodie/metadata` directory under +the basepath). In order to handle multi-writer scenario, users configure lock +provider and only one writer can access MDT in read-write mode. Hence, any write +to MDT is guarded by the data table lock. This ensures only one write is +committed to MDT at any point in time and thus guarantees serializability. +However, locking overhead adversely affects the write throughput and will reach +its scalability limits as we add more partitions to the MDT. + +## Goals + +- Support indexing one or more partitions in MDT while regular writers and table + services (such as cleaning or compaction) are in progress. +- Locking to be as lightweight as possible. +- Keep required config changes to a minimum to simplify deployment / upgrade in + production. +- Do not require specific ordering of how writers and table service pipelines + need to be upgraded / restarted. +- If an external long-running process is being used to initialize the index, the + process should be made idempotent so it can handle errors from previous runs. +- To re-initialize the index, make it as simple as running the external + initialization process again without having to change configs. + +## Implementation + +### A new Hudi action: INDEX + +We introduce a new action `index` which will denote the index building process, +the mechanics of which is as follows: + +1. From an external process, users can issue a CREATE INDEX or similar statement + to trigger indexing for an existing table. + 1. This will add a `<instant_time>.index.requested` to the timeline, which + contains the indexing plan. + 2. From here on, the index building process will continue to build an index + up to instant time `t`, where `t` is the latest completed instant time on + the timeline without any + "holes" i.e. no pending async operations prior to it. + 3. The indexing process will write these out as base files within the + corresponding metadata partition. A metadata partition cannot be used if + there is any pending indexing action against it. + +2. Any inflight writers (i.e. with instant time `t'` > `t`) will check for any + new indexing request on the timeline prior to preparing to commit. + 1. Such writers will proceed to additionally add log entries corresponding + to each such indexing request into the metadata partition. + 2. There is always a TOCTOU issue here, where the inflight writer may not + see an indexing request that was just added and proceed to commit without + that. We will correct this during indexing action completion. In the + average case, this may not happen and the design has liveness. + +3. When the indexing process is about to complete, it will check for all + completed commit actions to ensure each of them added entries per its + indexing plan, otherwise simply abort after a configurable timeout. Let's + call this the **indexing check**. Review comment: That's correct and if the indexer times out before C5 completes then it will abort. Next time, indexing will start again with C4 as base instant and run indexing check again. -- 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]
