decentraland / worlds-content-server / 27793704131

89%

main: 89%

Pull #496

github

fix: prevent RPC premature-close from crashing the process

A FetchError "Invalid response body ...: Premature close"
(ERR_STREAM_PREMATURE_CLOSE) from the RPC endpoint was taking down the
whole server.

Root cause: eth-connect's HTTPProvider consumes the response body with
`await res.json()` inside its `fetch(...).then(onFulfilled)` callback. When
the upstream RPC closes the connection mid-body, node-fetch rejects that
body read. Because the rejection happens in the fulfillment handler (not the
`.then(_, onRejected)` handler) and is never `.catch()`ed, it escapes as an
unhandled promise rejection. Under `--unhandled-rejections=strict` that
terminates the process. It also bypassed the retry logic in name-ownership,
which only sees errors delivered to the sendAsync callback.

Changes:
- Add createEthereumProvider, a fetch wrapper that reads and parses the RPC
  body inside the awaited fetch call. A premature close / stream error now
  rejects the fetch promise that HTTPProvider does handle, so it reaches the
  sendAsync callback as an ordinary, retryable error instead of crashing.
- Broaden sendBatchWithRetry to retry transient network failures (premature
  close, ECONNRESET, timeouts, DNS blips), not just HTTP 400s.
- Add a process-level unhandledRejection backstop that logs and keeps the
  server alive, and drop --unhandled-rejections=strict (the handler is a
  no-op while strict is set) so the backstop is effective.

1400 of 1678 branches covered (83.43%)

Branch coverage included in aggregate %.

10 of 21 new or added lines in 4 files covered. (47.62%)

2975 of 3235 relevant lines covered (91.96%)

68.04 hits per line