Skip to content

Limitations

What ankka-flow does not do, stated plainly so a design does not depend on it.

  • At least once, never exactly once. A record whose emits were written but whose offsets were not yet committed is delivered again after a failure or a rebalance. Streamlet logic must tolerate repeats. Delivery and failure explains when they happen.
  • JSON contracts only. Avro and Protobuf contracts, schema registries and compatibility rules between schema versions are not supported, and nothing checks a record against its schema. A new contract version is a new schema name.
  • Two SDKs. Scala and Python. Any other language can implement the streamlet protocol directly and prove itself with the descriptor fixtures and the conformance suite; see Adding a language SDK.
  • One built-in stage. The Neo4j merge sink is the only streamlet that runs inside the sidecar, and there is no way to add a stage from outside the sidecar image; every other streamlet's logic runs in its own process.
  • Neo4j 5.26 or later, written only. The merge sink needs dynamic labels in MERGE, which Neo4j 5.26 introduced; it refuses to open against an older server. No other graph database is supported, a pipeline cannot supply its own Cypher, and the sink never reads from the graph.
  • Graph deltas state, never change. A delta carries an element's whole state with a version from one source entity; there are no increments, no merging of properties from several sources, and tombstones mark elements rather than removing them. See Graph deltas.
  • The platform writes no delete markers. A tombstone stays in a delta topic as its element's last record; removing it from the topic is its writer's to do, and nothing removes a tombstoned element from the graph database.
  • A rebuilt graph is the live graph. A graph rebuilt from its delta topic has every element that is not marked deleted exactly as it was; tombstoned elements whose records have left the topic are not in it. See Rebuild a graph from its delta topic.
  • No per-partition state in the process. A streamlet sees batches of whichever partitions are assigned to its pod, and the assignment changes on every rebalance. State that must survive belongs in a topic or an external store.
  • No dead-letter topic and no skipping. A batch that fails every time stalls its partition, indefinitely, by design. The stall is visible as lag, a metric and a PartitionStalled event; skipping a record is the streamlet's own decision.
  • No HTTP or gRPC ingress into a pipeline. Records enter a pipeline through a Kafka topic.
  • Kafka only. No other broker, and no single topic spread over several Kafka clusters.
  • Existing topics are never changed, their cleanup policy included. The operator creates a managed topic once; a different partition count, replication or topic configuration on an existing topic is reported, not applied.
  • No user interface and no hosted control plane. A pipeline is a Kubernetes resource, operated with kubectl and the flow CLI.
  • Four platforms for the native CLI. flow ships as a native binary for macOS (Apple silicon and Intel) and Linux (x64 and arm64), through Homebrew or a release archive. Anywhere else it is built from source and runs on a JVM. There is no Windows build and no apt, dnf or winget package.
  • Records under 4 MiB. A record larger than the protocol's per-record limit, just under 4 MiB including its key and headers, fails the stream and stalls its partition.