The bar: state as a materialized view over the runtime own ordered log plus checkpoints — a separate durable event log is not required.
Two runtimes clear it. agentic-pekko uses real Pekko Persistence (EventSourcedBehavior + journal) and has a test proving the transcript survives entity restart while brain invocations stay flat. agentic-clj uses real Datomic with working as-of time travel.
The flagship Flink runtime does not. enableCheckpointing, CheckpointingMode, setStateBackend and CheckpointedFunction appear zero times in the framework — only in a build-excluded demo, and there with the non-durable HashMapStateBackend.
Supporting gaps:
PollingSource splits are explicitly stateless (serialize() returns new byte[0]), so every non-Kafka channel is at-most-once across a restart.- No sink implements
SupportsCommitter/TwoPhaseCommittingSink.
- Savepoint restore of an A2A bridge stream NPEs (snapshot restored with a null serializer arg).
- Both Flink-state vector memories call
getMapState() from non-keyed ProcessFunctions, so they throw in open() and are unreachable through the shipped corpus/retrieval API.
Deliverable: checkpointing configured (or explicitly documented as a caller responsibility), a durable state backend option, sources that actually snapshot offsets, and a test that restarts from a checkpoint and asserts the conversation view is intact.
Roadmap epic. Detailed findings, file:line citations and sub-tasks live in the AI-grind tracker (project AGS), which is the source of truth for this work. Local dev — build, test, tracking, profiling — is managed with AI-grind.
Tracked as AGS-6 in the devtools-mcp tracker.
The bar: state as a materialized view over the runtime own ordered log plus checkpoints — a separate durable event log is not required.
Two runtimes clear it.
agentic-pekkouses real Pekko Persistence (EventSourcedBehavior+ journal) and has a test proving the transcript survives entity restart while brain invocations stay flat.agentic-cljuses real Datomic with workingas-oftime travel.The flagship Flink runtime does not.
enableCheckpointing,CheckpointingMode,setStateBackendandCheckpointedFunctionappear zero times in the framework — only in a build-excluded demo, and there with the non-durableHashMapStateBackend.Supporting gaps:
PollingSourcesplits are explicitly stateless (serialize()returnsnew byte[0]), so every non-Kafka channel is at-most-once across a restart.SupportsCommitter/TwoPhaseCommittingSink.getMapState()from non-keyedProcessFunctions, so they throw inopen()and are unreachable through the shipped corpus/retrieval API.Deliverable: checkpointing configured (or explicitly documented as a caller responsibility), a durable state backend option, sources that actually snapshot offsets, and a test that restarts from a checkpoint and asserts the conversation view is intact.
Roadmap epic. Detailed findings, file:line citations and sub-tasks live in the AI-grind tracker (project
AGS), which is the source of truth for this work. Local dev — build, test, tracking, profiling — is managed with AI-grind.Tracked as
AGS-6in the devtools-mcp tracker.