async def _send_request(
self,
request: MatrixFederationRequest,
retry_on_dns_fail: bool = True,
timeout: Optional[int] = None,
long_retries: bool = False,
ignore_backoff: bool = False,
backoff_on_404: bool = False,
) -> IResponse:
"""
Sends a request to the given server.
Args:
request: details of request to be sent
retry_on_dns_fail: true if the request should be retied on DNS failures
timeout: number of milliseconds to wait for the response headers
(including connecting to the server), *for each attempt*.
60s by default.
long_retries: whether to use the long retry algorithm.
The regular retry algorithm makes 4 attempts, with intervals
[0.5s, 1s, 2s].
The long retry algorithm makes 11 attempts, with intervals
[4s, 16s, 60s, 60s, ...]
Both algorithms add -20%/+40% jitter to the retry intervals.
Note that the above intervals are *in addition* to the time spent
waiting for the request to complete (up to `timeout` ms).
NB: the long retry algorithm takes over 20 minutes to complete, with
a default timeout of 60s!
ignore_backoff: true to ignore the historical backoff data
and try the request anyway.
backoff_on_404: Back off if we get a 404
Returns:
Resolves with the HTTP response object on success.
Raises:
HttpResponseException: If we get an HTTP response code >= 300
(except 429).
NotRetryingDestination: If we are not yet ready to retry this
server.
FederationDeniedError: If this destination is not on our
federation whitelist
RequestSendFailed: If there were problems connecting to the
remote, due to e.g. DNS failures, connection timeouts etc.
"""
if timeout:
_sec_timeout = timeout / 1000
else:
_sec_timeout = self.default_timeout
if (self.hs.config.federation.federation_domain_whitelist is not None
and request.destination
not in self.hs.config.federation.federation_domain_whitelist):
raise FederationDeniedError(request.destination)
limiter = await synapse.util.retryutils.get_retry_limiter(
request.destination,
self.clock,
self._store,
backoff_on_404=backoff_on_404,
ignore_backoff=ignore_backoff,
)
method_bytes = request.method.encode("ascii")
destination_bytes = request.destination.encode("ascii")
path_bytes = request.path.encode("ascii")
if request.query:
query_bytes = encode_query_args(request.query)
else:
query_bytes = b""
scope = start_active_span(
"outgoing-federation-request",
tags={
tags.SPAN_KIND: tags.SPAN_KIND_RPC_CLIENT,
tags.PEER_ADDRESS: request.destination,
tags.HTTP_METHOD: request.method,
tags.HTTP_URL: request.path,
},
finish_on_close=True,
)
# Inject the span into the headers
headers_dict: Dict[bytes, List[bytes]] = {}
opentracing.inject_header_dict(headers_dict, request.destination)
headers_dict[b"User-Agent"] = [self.version_string_bytes]
with limiter, scope:
# XXX: Would be much nicer to retry only at the transaction-layer
# (once we have reliable transactions in place)
if long_retries:
retries_left = MAX_LONG_RETRIES
else:
retries_left = MAX_SHORT_RETRIES
url_bytes = request.uri
url_str = url_bytes.decode("ascii")
url_to_sign_bytes = urllib.parse.urlunparse(
(b"", b"", path_bytes, None, query_bytes, b""))
while True:
try:
json = request.get_json()
if json:
headers_dict[b"Content-Type"] = [b"application/json"]
auth_headers = self.build_auth_headers(
destination_bytes, method_bytes, url_to_sign_bytes,
json)
data = encode_canonical_json(json)
producer: Optional[
IBodyProducer] = QuieterFileBodyProducer(
BytesIO(data), cooperator=self._cooperator)
else:
producer = None
auth_headers = self.build_auth_headers(
destination_bytes, method_bytes, url_to_sign_bytes)
headers_dict[b"Authorization"] = auth_headers
logger.debug(
"{%s} [%s] Sending request: %s %s; timeout %fs",
request.txn_id,
request.destination,
request.method,
url_str,
_sec_timeout,
)
outgoing_requests_counter.labels(request.method).inc()
try:
with Measure(self.clock, "outbound_request"):
# we don't want all the fancy cookie and redirect handling
# that treq.request gives: just use the raw Agent.
# To preserve the logging context, the timeout is treated
# in a similar way to `defer.gatherResults`:
# * Each logging context-preserving fork is wrapped in
# `run_in_background`. In this case there is only one,
# since the timeout fork is not logging-context aware.
# * The `Deferred` that joins the forks back together is
# wrapped in `make_deferred_yieldable` to restore the
# logging context regardless of the path taken.
request_deferred = run_in_background(
self.agent.request,
method_bytes,
url_bytes,
headers=Headers(headers_dict),
bodyProducer=producer,
)
request_deferred = timeout_deferred(
request_deferred,
timeout=_sec_timeout,
reactor=self.reactor,
)
response = await make_deferred_yieldable(
request_deferred)
except DNSLookupError as e:
raise RequestSendFailed(
e, can_retry=retry_on_dns_fail) from e
except Exception as e:
raise RequestSendFailed(e, can_retry=True) from e
incoming_responses_counter.labels(request.method,
response.code).inc()
set_tag(tags.HTTP_STATUS_CODE, response.code)
response_phrase = response.phrase.decode("ascii",
errors="replace")
if 200 <= response.code < 300:
logger.debug(
"{%s} [%s] Got response headers: %d %s",
request.txn_id,
request.destination,
response.code,
response_phrase,
)
pass
else:
logger.info(
"{%s} [%s] Got response headers: %d %s",
request.txn_id,
request.destination,
response.code,
response_phrase,
)
# :'(
# Update transactions table?
d = treq.content(response)
d = timeout_deferred(d,
timeout=_sec_timeout,
reactor=self.reactor)
try:
body = await make_deferred_yieldable(d)
except Exception as e:
# Eh, we're already going to raise an exception so lets
# ignore if this fails.
logger.warning(
"{%s} [%s] Failed to get error response: %s %s: %s",
request.txn_id,
request.destination,
request.method,
url_str,
_flatten_response_never_received(e),
)
body = None
exc = HttpResponseException(response.code,
response_phrase, body)
# Retry if the error is a 5xx or a 429 (Too Many
# Requests), otherwise just raise a standard
# `HttpResponseException`
if 500 <= response.code < 600 or response.code == 429:
raise RequestSendFailed(exc,
can_retry=True) from exc
else:
raise exc
break
except RequestSendFailed as e:
logger.info(
"{%s} [%s] Request failed: %s %s: %s",
request.txn_id,
request.destination,
request.method,
url_str,
_flatten_response_never_received(e.inner_exception),
)
if not e.can_retry:
raise
if retries_left and not timeout:
if long_retries:
delay = 4**(MAX_LONG_RETRIES + 1 - retries_left)
delay = min(delay, 60)
delay *= random.uniform(0.8, 1.4)
else:
delay = 0.5 * 2**(MAX_SHORT_RETRIES - retries_left)
delay = min(delay, 2)
delay *= random.uniform(0.8, 1.4)
logger.debug(
"{%s} [%s] Waiting %ss before re-sending...",
request.txn_id,
request.destination,
delay,
)
await self.clock.sleep(delay)
retries_left -= 1
else:
raise
except Exception as e:
logger.warning(
"{%s} [%s] Request failed: %s %s: %s",
request.txn_id,
request.destination,
request.method,
url_str,
_flatten_response_never_received(e),
)
raise
return response
async def send_request(*, instance_name="master", **kwargs):
if instance_name == local_instance_name:
raise Exception("Trying to send HTTP request to self")
if instance_name == "master":
host = master_host
port = master_port
elif instance_name in instance_map:
host = instance_map[instance_name].host
port = instance_map[instance_name].port
else:
raise Exception("Instance %r not in 'instance_map' config" %
(instance_name, ))
data = await cls._serialize_payload(**kwargs)
url_args = [
urllib.parse.quote(kwargs[name], safe="")
for name in cls.PATH_ARGS
]
if cls.CACHE:
txn_id = random_string(10)
url_args.append(txn_id)
if cls.METHOD == "POST":
request_func = client.post_json_get_json
elif cls.METHOD == "PUT":
request_func = client.put_json
elif cls.METHOD == "GET":
request_func = client.get_json
else:
# We have already asserted in the constructor that a
# compatible was picked, but lets be paranoid.
raise Exception("Unknown METHOD on %s replication endpoint" %
(cls.NAME, ))
uri = "http://%s:%s/_synapse/replication/%s/%s" % (
host,
port,
cls.NAME,
"/".join(url_args),
)
try:
# We keep retrying the same request for timeouts. This is so that we
# have a good idea that the request has either succeeded or failed on
# the master, and so whether we should clean up or not.
while True:
headers: Dict[bytes, List[bytes]] = {}
# Add an authorization header, if configured.
if replication_secret:
headers[b"Authorization"] = [
b"Bearer " + replication_secret
]
opentracing.inject_header_dict(headers,
check_destination=False)
try:
result = await request_func(uri, data, headers=headers)
break
except RequestTimedOutError:
if not cls.RETRY_ON_TIMEOUT:
raise
logger.warning("%s request timed out; retrying", cls.NAME)
# If we timed out we probably don't need to worry about backing
# off too much, but lets just wait a little anyway.
await clock.sleep(1)
except HttpResponseException as e:
# We convert to SynapseError as we know that it was a SynapseError
# on the main process that we should send to the client. (And
# importantly, not stack traces everywhere)
_outgoing_request_counter.labels(cls.NAME, e.code).inc()
raise e.to_synapse_error()
except Exception as e:
_outgoing_request_counter.labels(cls.NAME, "ERR").inc()
raise SynapseError(502,
"Failed to talk to main process") from e
_outgoing_request_counter.labels(cls.NAME, 200).inc()
return result