fix: bound the error label of failed_store_queries (#4053)

This commit is contained in:
Darshan 2026-07-28 14:24:51 +05:30 committed by GitHub
parent 296460a407
commit 92841b5041
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -23,6 +23,22 @@ store_query_latency = Gauge('store_query_latency', 'Latency of the last store qu
consecutive_successful_responses = Gauge('consecutive_successful_responses', 'Consecutive successful store responses', ['node'])
node_health = Gauge('node_health', "Binary indicator of a node's health. 1 is healthy, 0 is not", ['node'])
# Bounded set of values for the 'error' label of failed_store_queries. A label must
# never hold a raw cause string: that is what blows up cardinality.
ERR_LOCAL_NODE_UNREACHABLE = 'local_node_unreachable'
ERR_PEER_UNREACHABLE = 'peer_unreachable'
ERR_QUERY_TIMEOUT = 'query_timeout'
ERR_STORE_BACKEND_ERROR = 'store_backend_error'
ERR_UNKNOWN = 'unknown'
# Store protocol status code -> error label. 300 (BAD_RESPONSE) is absent on purpose:
# it is the only status whose cause string has to be inspected to be classified.
STORE_STATUS_ERRORS = {
504: ERR_PEER_UNREACHABLE, # PEER_DIAL_FAILURE
503: ERR_STORE_BACKEND_ERROR, # SERVICE_UNAVAILABLE
429: ERR_STORE_BACKEND_ERROR, # TOO_MANY_REQUESTS: load shedding is store-side
}
# Argparser configuration
parser = argparse.ArgumentParser(description='')
@ -77,18 +93,49 @@ def send_sonda_msg(rest_address, pubsub_topic, content_topic, timestamp):
return False
# Maps the cause of a store BAD_RESPONSE (300) onto the bounded error set.
# This is the only place where a cause string is inspected.
def classify_bad_response(cause):
cause = cause.lower()
if 'archive error' in cause or 'driver_error' in cause:
return ERR_STORE_BACKEND_ERROR
if 'cancel' in cause or 'timed out' in cause or 'timeout' in cause:
return ERR_QUERY_TIMEOUT
if 'closed' in cause or 'reset' in cause or 'eof' in cause:
return ERR_PEER_UNREACHABLE
return ERR_UNKNOWN
# Increments failed_store_queries with a bounded 'error' label, keeping the raw cause in
# the log only. blame_peer=False leaves 'node' empty for failures where sonda never got
# an answer about the store peer, so we don't blame a peer that was never contacted.
def record_store_failure(store_node, error, cause, blame_peer=True):
log_with_utc(f'Failed store query node={store_node} error={error} cause={cause}')
failed_store_queries.labels(node=store_node if blame_peer else '', error=error).inc()
consecutive_successful_responses.labels(node=store_node).set(0)
# We return true if both our node and the queried Store node returned a 200
# If our message isn't found but we did get a store 200 response, this function still returns true
def check_store_response(json_response, store_node, timestamp):
status_code = json_response.get('statusCode')
status_desc = json_response.get('statusDesc') or ''
# Check for the store node status code
if json_response.get('statusCode') != 200:
error = f"{json_response.get('statusCode')} {json_response.get('statusDesc')}"
log_with_utc(f'Failed performing store query {error}')
failed_store_queries.labels(node=store_node, error=error).inc()
consecutive_successful_responses.labels(node=store_node).set(0)
if status_code != 200:
blame_peer = True
if status_code == 300:
error = classify_bad_response(status_desc)
elif status_code in STORE_STATUS_ERRORS:
error = STORE_STATUS_ERRORS[status_code]
else:
# A 400 means sonda built a bad query. That is a sonda bug, not the peer's fault.
error = ERR_UNKNOWN
blame_peer = False
record_store_failure(store_node, error, f'{status_code} {status_desc}', blame_peer)
return False
messages = json_response.get('messages')
# If there's no message in the response, increase counters and return
if not messages:
@ -132,27 +179,34 @@ def send_store_query(rest_address, store_node, encoded_pubsub_topic, encoded_con
try:
log_with_utc(f'Sending store request to {store_node}')
response = requests.get(url, params=params)
except requests.RequestException as e:
# No HTTP response at all, so sonda learned nothing about the store peer.
# Checked first: attributing this to a peer we never reached would be a lie.
record_store_failure(store_node, ERR_LOCAL_NODE_UNREACHABLE, str(e), blame_peer=False)
return False
except Exception as e:
log_with_utc(f'Error sending request: {e}')
failed_store_queries.labels(node=store_node, error=str(e)).inc()
consecutive_successful_responses.labels(node=store_node).set(0)
return False
# An exception inside sonda itself.
record_store_failure(store_node, ERR_UNKNOWN, str(e), blame_peer=False)
return False
elapsed_seconds = time.time() - s_time
log_with_utc(f'Response from {rest_address}: status:{response.status_code} [{elapsed_seconds:.4f} s.]')
# Our node turns a store TOO_MANY_REQUESTS into an HTTP 429 instead of a status code.
if response.status_code == 429:
record_store_failure(store_node, ERR_STORE_BACKEND_ERROR, response.text)
return False
if response.status_code != 200:
failed_store_queries.labels(node=store_node, error=f'{response.status_code} {response.content}').inc()
consecutive_successful_responses.labels(node=store_node).set(0)
record_store_failure(store_node, ERR_UNKNOWN,
f'HTTP {response.status_code} {response.text}', blame_peer=False)
return False
# Parse REST response into JSON
try:
json_response = response.json()
except Exception as e:
log_with_utc(f'Error parsing response JSON: {e}')
failed_store_queries.labels(node=store_node, error="JSON parse error").inc()
consecutive_successful_responses.labels(node=store_node).set(0)
except ValueError as e:
record_store_failure(store_node, ERR_UNKNOWN, str(e), blame_peer=False)
return False
# Analyze Store response. Return false if response is incorrect or has an error status