[benface]

On long-lived streams, Amp sends a keep-alive every keep_alive_interval (30 seconds by default): an empty FlightData with no data_header. streamQuery read every frame's header as a flatbuffer, so the first keep-alive killed the stream. It now skips frames with an empty data_header, the same check Amp's Rust client (amp-client-flight) makes in decode_flight_data.

• streamQuery with stream: true, and so ProtocolStream, TransactionalStream and CdcStream, died 30 seconds in with a defect: RangeError: Offset is outside the bounds of the DataView. Busy streams too, since Amp sends the keep-alives on a fixed interval.

I reproduced it with ampd v0.0.42 (ampd solo) over an Anvil chain with 2-second blocks: streamQuery('SELECT block_num FROM "_/anvil".blocks', { stream: true }) got a frame every 2 seconds, then died at 30 seconds with that RangeError. With the fix, it streamed past both keep-alives (30 and 60 seconds) with no gaps in the block numbers.

You might prefer checking dataBody too (dataHeader.length === 0 && dataBody.length === 0). I only check the header, like the Rust client.

Checks:

• pnpm lint, pnpm check:tsgo, pnpm test (234 passed, 1 new) and pnpm circular pass.
• The new test puts keep-alive frames before the schema, between the schema and a record batch, and after it. Without the fix it fails with the same RangeError.
• Against ampd v0.0.42, the stream that died at 30 seconds was still going at 75 seconds.

See edgeandnode/amp#2661 for the keep-alive frames on Amp's side.

🤖 Generated with Claude Code