The delivery workers use a lot of CPU, we could find out where they are

using lots of CPU by tracing the application to gather some data

Alternatively we could take a stab in the dark, which is what this
commit is doing.

I have the hypothesis that we are not re-using TCP connections using
HTTP keepalive

Refer to https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Keep-Alive

This means we are renegotiating the TLS connection every time we want to
send an SMS.  When we are sending lots of SMS messages then this will do
a lot of crypto handshaking which is expensive (in terms of CPU)

ie send_sms calls request() which creates a new tcp connection and a
new TLS handshake

When you use request.Session() to create a session it uses urllib3's
connection pooling, I've arbitrarily chosen 32 connections per pool
(with a default number of pools = 10)

ie init_app creates a session which has underlying connection pools
send_sms claims a connection from the pool and uses it to create or
re-use an existing TLS connection

Sessions are usually not great because they share data like cookies, but
when calling an API this is fine, or at least it is probably worth
canarying

Another way of re-using connections is by running HAProxy or similar as
a side-car proxy, which proxies to the API. send_sms would make a
local TCP connection to HAProxy which proxies to the MMG or Firetext
API via TLS, adding the Connection: keep-alive header

This command can be used to see how many TLS handshakes your computer is
doing, with some false positives:

tcpdump -n "tcp port 443 and (tcp[((tcp[12] & 0xf0) >> 2)] = 0x16)"

(alternatively we could just instrument the code /shrug)

Signed-off-by: toby lorne <toby@toby.codes>

app/clients/email/aws_ses.py

@@ -62,6 +62,7 @@ class AwsSesClient(EmailClient):
         # before-call, after-call, after-call-error, request-created, response-received
         self._client.meta.events.register('request-created.ses.SendEmail', self.ses_request_created_hook)
         self._client.meta.events.register('response-received.ses.SendEmail', self.ses_response_received_hook)
+        self._client.meta.events.register('before-call.ses', self.ses_inject_connection_header)
 
     def ses_request_created_hook(self, **kwargs):
         # request created may be called multiple times if the request auto-retries. We want to count all these as the
@@ -77,6 +78,11 @@ class AwsSesClient(EmailClient):
     def get_name(self):
         return self.name
 
+    def ses_inject_connection_header(self, params, **kwargs):
+        # keep underlying TLS connection open, so we do not spend lots of CPU
+        # and network time renegotiating TLS
+        params['headers']['Connection'] = 'Keep-Alive'
+
     def send_email(self,
                    source,
                    to_addresses,
@@ -148,3 +154,4 @@ def punycode_encode_email(email_address):
     # only the hostname should ever be punycode encoded.
     local, hostname = email_address.split('@')
     return '{}@{}'.format(local, hostname.encode('idna').decode('utf-8'))
+

app/clients/sms/firetext.py

@@ -2,7 +2,8 @@ import json
 import logging
 
 from time import monotonic
-from requests import request, RequestException
+from requests import request, RequestException, Session
+from requests.adapters import HTTPAdapter
 
 from app.clients.sms import (SmsClient, SmsClientResponseException)
 
@@ -69,6 +70,9 @@ class FiretextClient(SmsClient):
         self.name = 'firetext'
         self.url = current_app.config.get('FIRETEXT_URL')
         self.statsd_client = statsd_client
+        # this uses urllib3 under the hood to create a connection pool
+        self.session = Session()
+        self.session.mount('https://', HTTPAdapter(pool_maxsize=32))
 
     def get_name(self):
         return self.name
@@ -103,8 +107,7 @@ class FiretextClient(SmsClient):
         response = None
         start_time = monotonic()
         try:
-            response = request(
-                "POST",
+            response = self.session.post(
                 self.url,
                 data=data,
                 timeout=60

app/clients/sms/mmg.py

@@ -1,6 +1,8 @@
 import json
 from time import monotonic
-from requests import (request, RequestException)
+from requests import (request, RequestException, Session)
+from requests.adapters import HTTPAdapter
+
 from app.clients.sms import (SmsClient, SmsClientResponseException)
 
 mmg_response_map = {
@@ -75,6 +77,9 @@ class MMGClient(SmsClient):
         self.name = 'mmg'
         self.statsd_client = statsd_client
         self.mmg_url = current_app.config.get('MMG_URL')
+        # this uses urllib3 under the hood to create a connection pool
+        self.session = Session()
+        self.session.mount('https://', HTTPAdapter(pool_maxsize=32))
 
     def record_outcome(self, success, response):
         status_code = response.status_code if response else 503
@@ -108,8 +113,8 @@ class MMGClient(SmsClient):
         response = None
         start_time = monotonic()
         try:
-            response = request(
-                "POST",
+
+            response = self.session.post(
                 self.mmg_url,
                 data=json.dumps(data),
                 headers={