Cool, thanks Ryan, John, Amogh for the reply! Great to see you interested! Felix will have a Spark Scalability & Reliability Sync meeting on Dec 4 1pm PST. We could discuss more details there. Do you want to join?On Tue, Nov 19, 2019 at 4:23 PM Amogh Margoor <email@example.com> wrote:We at Qubole are also looking at disaggregating shuffle on Spark. Would love to collaborate and share learnings.Regards,AmoghOn Tue, Nov 19, 2019 at 4:09 PM John Zhuge <firstname.lastname@example.org> wrote:Great work, Bo! Would love to hear the details.On Tue, Nov 19, 2019 at 4:05 PM Ryan Blue <email@example.com> wrote:I'm interested in remote shuffle services as well. I'd love to hear about what you're using in production!rbOn Tue, Nov 19, 2019 at 2:43 PM bo yang <firstname.lastname@example.org> wrote:Hi Ben,Thanks for the writing up! This is Bo from Uber. I am in Felix's team in Seattle, and working on disaggregated shuffle (we called it remote shuffle service, RSS, internally). We have put RSS into production for a while, and learned a lot during the work (tried quite a few techniques to improve the remote shuffle performance). We could share our learning with the community, and also would like to hear feedback/suggestions on how to further improve remote shuffle performance. We could chat more details if you or other people are interested.Best,BoOn Fri, Nov 15, 2019 at 4:10 PM Ben Sidhom <email@example.com> wrote:
I would like to start a conversation about extending the Spark shuffle manager surface to support fully disaggregated shuffle implementations. This is closely related to the work in SPARK-25299, which is focused on refactoring the shuffle manager API (and in particular, SortShuffleManager) to use a pluggable storage backend. The motivation for that SPIP is further enabling Spark on Kubernetes.
The motivation for this proposal is enabling full externalized (disaggregated) shuffle service implementations. (Facebook’s Cosco shuffle is one example of such a disaggregated shuffle service.) These changes allow the bulk of the shuffle to run in a remote service such that minimal state resides in executors and local disk spill is minimized. The net effect is increased job stability and performance improvements in certain scenarios. These changes should work well with or are complementary to SPARK-25299. Some or all points may be merged into that issue as appropriate.
Below is a description of each component of this proposal. These changes can ideally be introduced incrementally. I would like to gather feedback and gauge interest from others in the community to collaborate on this. There are likely more points that would be useful to disaggregated shuffle services. We can outline a more concrete plan after gathering enough input. A working session could help us kick off this joint effort; maybe something in the mid-January to mid-February timeframe (depending on interest and availability. I’m happy to host at our Sunnyvale, CA offices.
Scheduling and re-executing tasks
Allow coordination between the service and the Spark DAG scheduler as to whether a given block/partition needs to be recomputed when a task fails or when shuffle block data cannot be read. Having such coordination is important, e.g., for suppressing recomputation after aborted executors or for forcing late recomputation if the service internally acts as a cache. One catchall solution is to have the shuffle manager provide an indication of whether shuffle data is external to executors (or nodes). Another option: allow the shuffle manager (likely on the driver) to be queried for the existence of shuffle data for a given executor ID (or perhaps map task, reduce task, etc). Note that this is at the level of data the scheduler is aware of (i.e., map/reduce partitions) rather than block IDs, which are internal details for some shuffle managers.
Add a heartbeat (keep-alive) mechanism to RDD shuffle output so that the service knows that data is still active. This is one way to enable time-/job-scoped data because a disaggregated shuffle service cannot rely on robust communication with Spark and in general has a distinct lifecycle from the Spark deployment(s) it talks to. This would likely take the form of a callback on ShuffleManager itself, but there are other approaches.
Add lifecycle hooks to shuffle readers and writers (e.g., to close/recycle connections/streams/file handles as well as provide commit semantics). SPARK-25299 adds commit semantics to the internal data storage layer, but this is applicable to all shuffle managers at a higher level and should apply equally to the ShuffleWriter.
Do not require ShuffleManagers to expose ShuffleBlockResolvers where they are not needed. Ideally, this would be an implementation detail of the shuffle manager itself. If there is substantial overlap between the SortShuffleManager and other implementations, then the storage details can be abstracted at the appropriate level. (SPARK-25299 does not currently change this.)
Do not require MapStatus to include blockmanager IDs where they are not relevant. This is captured by ShuffleBlockInfo including an optional BlockManagerId in SPARK-25299. However, this change should be lifted to the MapStatus level so that it applies to all ShuffleManagers. Alternatively, use a more general data-location abstraction than BlockManagerId. This gives the shuffle manager more flexibility and the scheduler more information with respect to data residence.
Allow serializers to be used more flexibly and efficiently. For example, have serializers support writing an arbitrary number of objects into an existing OutputStream or ByteBuffer. This enables objects to be serialized to direct buffers where doing so makes sense. More importantly, it allows arbitrary metadata/framing data to be wrapped around individual objects cheaply. Right now, that’s only possible at the stream level. (There are hacks around this, but this would enable more idiomatic use in efficient shuffle implementations.)
Have serializers indicate whether they are deterministic. This provides much of the value of a shuffle service because it means that reducers do not need to spill to disk when reading/merging/combining inputs--the data can be grouped by the service, even without the service understanding data types or byte representations. Alternative (less preferable since it would break Java serialization, for example): require all serializers to be deterministic.
- Ben--Ryan BlueSoftware EngineerNetflix--John Zhuge