HDDS-15864. Branch merge checklist for Short Circuit read #503
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,113 @@ | ||
| # HDDS-10685 : Short-Circuit Read | ||
|
|
||
| Epic: [HDDS-10685](https://issues.apache.org/jira/browse/HDDS-10685) | ||
| Feature branch: https://github.com/apache/ozone/tree/HDDS-10685 | ||
|
|
||
| ## 1. Builds/intermittent test failures | ||
|
|
||
| There are no intermittent failures specific to the HDDS-10685 branch as of now. During the development, it was ensured all the CI checks were clean prior to every commit merge. | ||
|
|
||
| The plan is to run repeated CI checks on the merge commit to master. | ||
|
|
||
| ## 2. Documentation | ||
|
|
||
| [User Documentation](https://ozone.apache.org/docs/next/administrator-guide/configuration/performance/short-circuit-local-reads/) of Short Circuit Read has been added. | ||
|
|
||
| ## 3. Design, attached the docs | ||
|
|
||
| Design document can be found here : [Short Circuit Read Support](https://github.com/apache/ozone/blob/HDDS-10685/hadoop-hdds/docs/content/design/short-circuit-read.md). | ||
|
|
||
| ## 4. S3 compatibility | ||
|
|
||
| N/A, S3 compatibility remains the same. Short Circuit Read only affects the client and DataNode read path. | ||
|
|
||
| ## 5. Docker-compose / Acceptance tests | ||
|
|
||
| New robot test [short-circuit.robot](https://github.com/apache/ozone/blob/HDDS-10685/hadoop-ozone/dist/src/main/smoketest/short-circuit/short-circuit.robot) is being added. | ||
|
|
||
| New acceptance tests are added, mainly tests the Short Circuit Read metrics. It does not test fault injection. | ||
|
|
||
| ## 6. Support of containers / Kubernetes | ||
|
|
||
| No addition. No change in existing support. | ||
|
|
||
| ## 7. Coverage / Code quality | ||
|
|
||
| [New Code Coverage](https://sonarcloud.io/summary/new_code?id=hadoop-ozone&branch=HDDS-10685) for Short Circuit Read Support (HDDS-10685) is 80.78%% and [Overall Code Coverage](https://sonarcloud.io/summary/overall?id=hadoop-ozone&branch=HDDS-10685) is 78.5%. | ||
| [Overall Code Coverage](https://sonarcloud.io/summary/overall?id=hadoop-ozone&branch=master) for master is 78.5%%. | ||
|
|
||
| ## 8. Build time | ||
|
|
||
| [Build time for the latest commit](https://github.com/apache/ozone/actions/runs/27126447271/job/80056268213) from HDDS-10685 Branch is 9m 59s. | ||
| [Build time for the latest commit](https://github.com/apache/ozone/actions/runs/29915076929/job/88926295859) from the master branch is 10m 43s. | ||
|
|
||
| ## 9. Possible incompatible changes/used feature flag | ||
|
|
||
| Short-circuit read is gated by the non-rolling upgrade framework as HDDSLayoutFeature.SHORT_CIRCUIT_READS (layout version 11). | ||
|
Contributor
There was a problem hiding this comment. Why is the layout feature necessary? It looks like this just creates a socket and closes it on shutdown. How would this break on downgrade? Also it looks like new clients can request short circuit read from an old server it will be ignored. This seems fine since the field is phrased as "request" but that is worth noting in this section too.
Contributor
Author
There was a problem hiding this comment. It's to avoid a new client talk to an old server. Yes, the new client has fallback mechanism, but it will cost time.
Contributor
There was a problem hiding this comment. Then this is not correct and was not tested. It is always the newest component's responsibility to handle compatibility, and Datanode layout feature is never communicated to the client so it is impossible for a new client to do anything about this with the current implementation. If we actually want the new client to do some handling with an old server, the layout feature should be removed and a The two version system is confusing, that's why we've reduced it to a single version in ZDU so this type of thing should be easier to follow going forward. |
||
|
|
||
| It remains inactive until the cluster upgrade is finalized. A global enable/disable switch is provided via ozone.client.read.short-circuit (default: false). | ||
|
|
||
| To enable the feature (after finalization), add the following to the client and Datanode ozone-site.xml to enable the feature: | ||
|
|
||
| ```xml | ||
| <property> | ||
| <name>ozone.client.read.short-circuit</name> | ||
| <value>true</value> | ||
| <tag>CLIENT, DATANODE</tag> | ||
| <description>Disable or enable the short-circuit local read feature. By default it is disabled.</description> | ||
| </property> | ||
| ``` | ||
| And the following to the client and Datanode ozone-site.xml, to specify the path of the UNIX domain socket: | ||
| ```xml | ||
| <property> | ||
| <name>ozone.domain.socket.path</name> | ||
| <value></value> | ||
| <tag>CLIENT, DATANODE</tag> | ||
| <description>UNIX domain socket path for co-located client–Datanode short-circuit communication.</description> | ||
| </property> | ||
| ``` | ||
|
|
||
| ## 10. Third-party dependencies/License changes | ||
|
|
||
| There are no third party dependencies introduced by this feature. | ||
|
|
||
| ## 11. Performance | ||
|
|
||
| The major workflow of Short-Circuit Read is: The client detects a local Datanode replica → opens a UNIX domain socket → sends GetBlock with requestShortCircuitAccess=true → Datanode passes a file descriptor → The client reads block data directly from the local disk via FileChannel. | ||
|
|
||
| End-to-end read performance is therefore dominated by local disk I/O, similar to a direct file read, rather than by gRPC data transfer. | ||
|
|
||
| A benchmark was run against feature branch HDDS-10685, testing short-circuit performance with the `ozone fs` command and YCSB on a ycloud cluster. Short-circuit read was toggled with `ozone.client.read.short-circuit` and `ozone.domain.socket.path`. | ||
|
|
||
| ### Cluster configuration | ||
|
|
||
|  | ||
|
|
||
| ### Ozone FS | ||
|
|
||
| Use `ozone fs -get ofs://ozone1733996033/vol-scr/buck/scr/file33 ./file33` to download a 10 GB file. Node9 and node7 have the datanode role; node8 does not. | ||
|
|
||
|  | ||
|
|
||
| ### YCSB | ||
|
|
||
| The tests are performed by running 3 consecutive iterations after changing the `ozone.client.read.short-circuit` configuration and restarting all related services. HBase `l1CacheHitRatio` is around 90% during the test. | ||
|
|
||
| **Workload C** | ||
|
|
||
|  | ||
|
|
||
| **Workload A** | ||
|
|
||
|  | ||
|
|
||
| Metrics (`ContainerLocalOps`, local op latencies, local bytes stats) are added in Datanode to observe the runtime state/latency. | ||
|
|
||
| ## 12. Security considerations | ||
|
|
||
| Short-Circuit Read does not introduce any new CLI or admin command. | ||
|
|
||
| Short-circuit communication uses a UNIX domain socket (`ozone.domain.socket.path`) between the client and Datanode. | ||
|
|
||
| It follows the same [Hadoop Socket Path Security](https://cwiki.apache.org/confluence/spaces/HADOOP2/pages/120730260/SocketPathSecurity) rules as HDFS short-circuit reads. | ||
|
Contributor
Contributor
There was a problem hiding this comment. agreed, as a TODO item we should copy this to our doc site.
Contributor
Author
There was a problem hiding this comment. https://cwiki.apache.org/confluence/spaces/HADOOP2/pages/120730260/SocketPathSecurity content is included in the design document "Security" section.
Contributor
There was a problem hiding this comment. Let's update this section to reference our own user or design docs with this content instead of linking to the hadoop wiki. |
||
|
|
||
Uh oh!
There was an error while loading. Please reload this page.