Fix json decode error

[npow]

Problem

Intermittent JSONDecodeError when multiple environments are resolved concurrently during deployment:

json.decoder.JSONDecodeError: Expecting value: line 1 column 86673 (char 86672)

Root Cause

Race condition in FIFO-based IPC between deployer subprocess and parent process:

1. Writer side: Subprocess writes JSON to FIFO, but Python's buffered I/O may not flush immediately
2. Reader side: Parent process reads from FIFO in non-blocking mode
3. Race: When subprocess exits quickly after close(), reader detects process exit and breaks on empty read
4. Problem: OS kernel may still have buffered data in pipe that hasn't been delivered yet
5. Result: Truncated JSON at arbitrary positions (~86KB in the error case)

Solution

Changed read_from_fifo_when_ready() to use a hybrid approach:

1. Start in non-blocking mode (existing behavior)

• Use select.poll() to wait for data
• Can detect subprocess failures early
• Can timeout if subprocess hangs

2. Switch to blocking mode once first data arrives

• Use fcntl() to remove O_NONBLOCK flag
• Continue with blocking read() calls
• POSIX guarantee: Blocking read() returns EOF (0 bytes) ONLY after writer closes AND all kernel pipe buffers are drained

[nflx-mf-bot]

Netflix internal testing[1398] @ 7594e0f

[romain-intel]

When does this happen? So we got some event (like file close?) and no data?

[npow]

If poll() returned an event (any event) AND read() returned 0 bytes then it must be EOF:

• If it was POLLIN (data ready), then read() would have returned data
• If read() returned 0 bytes despite an event, it must be POLLHUP (writer closed) or a stale POLLIN that resolved to EOF

[aquarin]

This is a very unit-testable chunk of code. Can we add unit test to:

1. To trigger the scenario that caused the error.
2. Verify the unit test fails before this code change.
3. Verify the unit test passes after this fix.

[npow]

We could add unit tests for this chunk of code, but the actual bug is difficult to reproduce reliably because of a race condition: the writer has to close while kernel buffers still has data