[Tracking Umbrella] Prism Runner areas for contribution.

[lostluck]

What needs to happen?

This issue is to track and refer to other issues/prs for various prism features. This issue shouldn't generally be commented on, but have this top entry edited as needed, referring to granular issues for individual features and support.

Ultimately, this will eventually track support in the Beam Compatibility Matrix, and keeping the Prism README up to date.

Complete items should be checked, and have links to their completing PR or closed primary tracking issue.

Items marked should only have an issue filed when the work has started, typically there's a meaningful design proposal, and understanding of what the closing criteria are. This can be "X set of existing SDK tests now pass", or a given capability is possible (eg. UI related features.)

Prism Areas for Contribution

Beam Core Priorities

These are features that prevent Prism use and adoption.

In progress by @lostluck

• State Handling ([Task][prism]: Support State API  #28543)
• [prism] Timer Handling (Event Time) #29772
  • Event Time
  • [prism] Processing Time handling (TestStream, Timers, ProcessContinuations) #30083
    • Using Real Time, when operating in "production" mode.
  • Timer Clearing
  • OnWindowExpiry
• [prism] TestStream support (Elements + Watermarks) #29917
  • Watermark and Element Event handling
  • [prism] Processing Time handling (TestStream, Timers, ProcessContinuations) #30083
• [prism] Triggers and Panes and Windowing Strategies #31438

Other Beam Core

This is an incomplete list of Beam features that would be nice to have.

• Multi-Chunk Iterable protocol ([Feature Request][prism]: Add support for the multi-chunk iterable protocol #27762)
• State Backed Iterables (for GBKs & Side Inputs)
  • State based Continuation Tokens
• OrderedListState
  • Blocked on inclusion in the FnAPI once in the FnAPI

Non-Go Blockers

Notable issues found in trying to run the Non Go SDKs (Java, Python, or others). Tracked in #28187, and more granular issues should be referred to here.

• Go SDK Cross Language PreCommit Suite ([#28187][prism] Basic cross language support. #28545)
  • No Basic Xlang Tests filtered.
• Prism Java Validates Runner Suite
  • Executing targets exists
  • No tests filtered.
• Python Validates Runner Suite
  • Executing targets exists
  • No tests filtered.
• Properly respect and handle SDK & Runner Capabilities.

Persistence & Reliability Features

Prism currently stores everything in memory. This includes all element data, in progress bundle data, pipeline info, artifacts etc. This is fast, but not the best use of memory for using prism long term as a stand alone runner.

• Per Pipeline data should be moved to a local file cache.
  • They aren’t stored in memory when not needed. Eg. Artifacts shouldn’t live in memory once necessary environments are spun up.
  • Garbage collect artifacts after pipeline termination.
  • Garbage collect older pipelines after some threshold.
• Pipeline Restarts
  • Optimized stages need to be stored, so no complex mapping needs to occur for any persisted state.
  • Per stage pending elements and state needs to be stored so bundles can be re-computed on restarts.
    • It should be possible for a pipeline to be aborted, and prism torn down, and for a pipeline to be restarted from where it left off, with new worker processes.
    • FrostDB is an embeddable-in-Go, write optimized, in-memory + persistence, columnar database that might be a good thing to look at to enable these features.
• Bundles Retries
  • Prism currently doesn’t retry failed bundles. A bundle failure fails the pipeline.
  • Adding a sensible retries policy would improve bundle reliability.
  • Affects how elements are divided into bundles, and scheduled.
  • Eg. A failed bundle could be split into smaller and smaller bundles, until the failing elements are isolated. Such a strategy would also enable implementation of error tolerance policies for example.
• Improve Bundle Splitting
  • Prism currently schedules all available pending elements into a single bundle.
  • Instead it could use some heuristic to determine how to split pending elements into new bundles to improve worker level parallelism before Channel or Sub Element Splitting occurs.
• [prism] Programmatic Cancel, and Drain #29669
• Pipeline Update

• Similar to Cancel + Drain in combination with Pipeline restarts. Allow a pipeline to be updated mid execution.

• Use the worker-id GRPC metadata
  • Distinguishes between pipeline workers to avoid needing a new port for each instance.
  • Need a single "multiplexer" layer to route between the handlers for given jobs and workers.

Performance features

These are non-user facing Beam features that Dataflow implements. In order for Prism to serve the purpose of validating pipeline locally before production runner execution, these are required, to reduce worker side execution differences.

• Side Input + State Cache
• Elements on ProcessBundleRequest
• Elements on ProcessBundleResponse
• Autosharded keys
• Map Side Input Keys
• Eagerly Bucket elements by Key + Window for GBKs
• Intern user-keys, tags strings, byte arrays to reduce memory bloat during stateful pipeline execution. Go is garbage collected, not magic.

Stand Alone UI Based Features

These are features that are best tied to the ability to understand a job in the UI.

• Data Sampling + plumbing to UI
• Worker Status support + plumbing to UI
• Runner side PubSub Transform (due to being a Beam built in)
• Display of Optimized stages in UI
• Display of Graph structure in UI
  • Interactivity with same.
• Display of Job Logs in UI
  • ...and storage thereof in local cache.

Other features

The following are known issues/desires without a specific categorization at present.

• Custom WindowFns
• Prism Per Job Configurability
  • Being able to toggle or set specific configurations using PipelineOptions or similar.
  • AKA the described Variants approach.
• Additional runner side execution metrics
  • Count Splits per Transform
  • Count executed "Bundles" Executed.
  • "stage" execution time
  • Num elements per stage
    • Num Keys per stage (if stateful)
  • Histograms or Timeseries of the above?

Completed Work

This section should be structured similarly to the Beam Compatibility Matrix for ease of transition to populating it there.

• Environment Execution
  • LOOPBACK/External
  • Docker
• Basics
  • DoFns
  • GBKS
  • Windows
  • Side Inputs
    • Map
    • Iterator
• Scaling
  • Splittable DoFn support
  • ProcessContinuation support
• Performance
  • Fusion