JSON workers should delimit work requests with newlines #14218
Description
Description of the problem / feature request:
JSON workers are difficult to write because work requests are not newline delimited. This issue was brought up in #13599, and the discussion generally indicated that this is a usability bug.
The proposal is to add a newline between requests according to ndjson.
Feature requests: what underlying problem are you trying to solve with this feature?
Right now an implementation of a JSON worker work request loop using python looks like this:
# This is a pretty cumbersome way of getting work requests.
# It waits for something to be available on stdin, and then tries parsing
# whatever text is available as JSON. It keeps accumulating more text until a
# full work request is available.
def wait_for_stdin_json():
stdin = open("/dev/stdin", "rb")
os.set_blocking(stdin.fileno(), False)
stream_bytes = b""
while True:
_, _, _ = select.select([stdin.fileno()], [], [])
maybe_bytes = stdin.read()
if maybe_bytes is None:
continue
stream_bytes += maybe_bytes
try:
work_request = json.loads(stream_bytes)
except json.decoder.JSONDecodeError:
# stream couldn't decode, wait until next select
sys.stderr.write(f"failed to decode: {stream_bytes.decode('utf-8')}\n")
sys.stderr.flush()
continue
yield work_request
stream_bytes = b"" # resetWith newline-delimited JSON it would look like this:
# Simplified json loop with newline-delimited json.
def wait_for_stdin_json():
while True:
yield json.loads(sys.stdin.readline())