Serverless in production an experience report
Yan Cui
What’s in this talk?
! how to responsibly run a serverless architecture (aka. how to do ops in serverless)
! testing, CI/CD! logging, distributed tracing, monitoring! config management, securing secrets! coldstarts! gotchas/limitations + workarounds/hacks
hi, I’m Yan Cui
hi, I’m Yan CuiAWS user since 2009
apr, 2016
Before
! hidden complexities and dependencies! low utilisation to leave headroom for large spikes! EC2 scaling is slow, so scale earlier! paying for lots of used resources! up to 30 mins to deploy! deployments required downtime
- Dan North
“lead time to someone saying thank you is the only reputation
metric that matters.”
“what would good look like for us?”
Deployments should…
! be small! be fast! have zero downtime! require no lock-step
Features should…
! be independently deployable! be loosely-coupled
We want to…
! minimise cost of unused resources! minimise ops effort! reduce technical mess! deliver visible improvements to users faster
nov, 2016
170 Lambda functions in prod
1.2 GB deployment packages in prod
95% cost saving vs EC2
15x no. of prod releases per month
timeis a good fit
1st function in prod!time
is a good fit
?
timeis a good fit
1st function in prod!
Practices ToolsPrinciples
what is good? how to make it good? with what?
Principles outlast Tools
ALERTING
CI / CD
TESTING
LOGGING
MONITORING
170 functions
WOOF!
? ?
timeis a good fit
1st function in prod!
CONFIG MANAGEMENT
SECURITY
DISTRIBUTED TRACING
evolving the platform
building a better search experience
Legacy Monolith Amazon Kinesis Amazon Lambda
Amazon CloudSearch
Legacy Monolith Amazon Kinesis Amazon Lambda
Amazon CloudSearchAmazon API Gateway Amazon Lambda
building an analytics pipeline
Legacy Monolith Amazon Kinesis Amazon Lambda
Google BigQuery
Legacy Monolith Amazon Kinesis Amazon Lambda
Google BigQuery1 developer, 2 daysdesign production
(his 1st serverless project)
Legacy Monolith Amazon Kinesis Amazon Lambda
Google BigQuery“thank you, nothing ever got
done this fast at Skype!”
- Dan North
“lead time to someone saying thank you is the only reputation
metric that matters.”
rebuilding the timeline feature
building better user recommendations
BigQuery
BigQuery
grapheneDB
BigQuery
grapheneDB
BigQuery
grapheneDB
BigQuery
getting PRODUCTION READY
CHOOSE A
FRAMEWORK
DEPLOYMENT
http://serverless.com
https://github.com/awslabs/serverless-application-model
http://apex.run
https://apex.github.io/up
https://github.com/claudiajs/claudia
https://github.com/Miserlou/Zappa
http://gosparta.io/
TESTING
amzn.to/29Lxuzu
Level of Testing
1.Unitdo our objects do the right thing?are they easy to work with?
1.Unit2.Integrationdoes our code work against code we can’t change?
Level of Testing
handler
handler
test by invoking the handler
Level of Testing
1.Unit2.Integration3.Acceptancedoes the whole system work?
Level of Testing
unit
integration
acceptance
feedb
ack
confidence
“…We find that tests that mock external libraries often need to be complex to get the code into the right state for the functionality we need to exercise.
The mess in such tests is telling us that the design isn’t right but, instead of fixing the problem by improving the code, we have to carry the extra complexity in both code and test…”
Don’t Mock Types You Can’t Change
“…The second risk is that we have to be sure that the behaviour we stub or mock matches what the external library will actually do…
Even if we get it right once, we have to make sure that the tests remain valid when we upgrade the libraries…”
Don’t Mock Types You Can’t Change
ServicesDon’t Mock Types You Can’t Change
Paul Johnston
The serverless approach to testing is different and
may actually be easier.http://bit.ly/2t5viwK
LambdaAPI Gateway DynamoDB
LambdaAPI Gateway DynamoDB
Unit Tests
LambdaAPI Gateway DynamoDB
Unit Tests
Mock/Stub
is our request correct?
is the request mapping set up
is the API resources configured correctly?
are we assuming the correct schema?
LambdaAPI Gateway DynamoDB
is Lambda proxy configured correctly?
is IAM policy set up correctly?
is the table created?
what unit tests will not tell you…
most Lambda functions are simple have single purpose, the risk of shipping broken
software has largely shifted to how they integrate with external services
observation
But it slows down my feedback loop…
IT’S NOT ABOUT YOU!
me
test your system, not (just) your code
API Gateway
IOT
Kinesis
SNS
ElastiCache
CloudWatch
DynamoDB
IAM
S3
Auth0
GrapheneDB
SES
Twilio
Google BigQuery
MongoLab
CloudSearch
APN
GCM
Lambda
EC2
…if a service can’t provide you with a relatively easy way to test the
interface in reality, then you should consider using another one.
Paul Johnston
“…Wherever possible, an acceptance test should exercise the system end-to-end without directly calling its internal code.
An end-to-end test interacts with the system only from the outside: through its interface…”
Testing End-to-End
Legacy Monolith Amazon Kinesis Amazon Lambda
Amazon CloudSearchAmazon API Gateway Amazon Lambda
Legacy Monolith Amazon Kinesis Amazon Lambda
Amazon CloudSearchAmazon API Gateway Amazon Lambda
Test Input
Legacy Monolith Amazon Kinesis Amazon Lambda
Amazon CloudSearchAmazon API Gateway Amazon Lambda
Test Input
Validate
integration tests exercise system’s Integration with its
external dependencies
acceptance tests exercise system End-to-End from
the outside
integration tests differ from acceptance tests only in HOW the
Lambda functions are invoked
observation
CI/CD PIPELINE
“…We prefer to have the end-to-end tests exercise both the system and the process by which it’s built and deployed…
This sounds like a lot of effort (it is), but has to be done anyway repeatedly during the software’s lifetime…”
Testing End-to-End
me
Deployment scripts that only live on the CI box is a disaster
waiting to happen.
Jenkins build config deploys and tests
unit + integration tests
deploy
acceptance tests
if [ "$1" = "deploy" ] && [ $# -eq 4 ]; then STAGE=$2 REGION=$3 PROFILE=$4
npm install AWS_PROFILE=$PROFILE 'node_modules/.bin/sls' deploy -s $STAGE -r $REGION …
if [ "$1" = "deploy" ] && [ $# -eq 4 ]; then STAGE=$2 REGION=$3 PROFILE=$4
npm install AWS_PROFILE=$PROFILE 'node_modules/.bin/sls' deploy -s $STAGE -r $REGION …
install serverless framework as dev dependency
can be run locally & on the CI box
auto auto manual
LOGGING
2016-07-12T12:24:37.571Z 994f18f9-482b-11e6-8668-53e4eab441ae GOT is off air, what do I do now?
2016-07-12T12:24:37.571Z 994f18f9-482b-11e6-8668-53e4eab441ae GOT is off air, what do I do now?
UTC Timestamp API Gateway Request Id
your log message
function name
date
function version
me
Logs are not easily searchable in CloudWatch Logs.
LOG OVERLOAD
CENTRALISE LOGS
CENTRALISE LOGS
MAKE THEM EASILYSEARCHABLE
+ +the elk stack
CloudWatch Logs
CloudWatch Logs AWS Lambda ELK stack
CloudWatch Events
CloudWatch Logs
http://bit.ly/2f3zxQG
DISTRIBUTED TRACING
“my followers didn’t receive my new post!”
- a user
where could the problem be?
correlation IDs*
* eg. request-id, user-id, yubl-id, etc.
ROLL YOUR OWNCLIENTS
kinesis client
http client
sns client
http://bit.ly/2k93hAj
kinesisglobal.CONTEXT
log.info(…)
api-b
global.CONTEXT
global.CONTEXT
global.CONTEXT
x-correlation-id = … x-correlation-xxx = …
API Gateway Kinesis
SNS
API Gateway
API Gatewayapi-a api-c
sns
headers[“User-Agent”] headers[“Debug-Log-Enabled”]
MessageAttributes: [ “x-correlation-id”: … “User-Agent”: … “Debug-Log-Enabled”: … ]
global.CONTEXT
headers[“User-Agent”] headers[“Debug-Log-Enabled”] headers[“x-correlation-id”]
headers[“User-Agent”] headers[“Debug-Log-Enabled”] headers[“x-correlation-id”]
data.__context
capture
forward
function
event
ROLL YOUR OWNCLIENTS
X-RAY
Amazon X-Ray
Amazon X-Ray
traces do not span over API Gateway
MONITORING + ALERTING
“where do I install monitoring agents?”
you can’t
• invocation Count• error Count• latency• throttling• granular to the minute• support custom metrics
• invocation Count• error Count• latency• throttling• granular to the minute• support custom metrics
Why not IOPipe?
! pervasive access to your entire application! adds latency for tracking
me
The only “background” processing you get are the capabilities the platform provides out of the box.
“how do I batch up and send logs/metrics in the
background?”
you can’t (kinda)
console.log(“hydrating yubls from db…”);
console.log(“fetching user info from user-api”);
console.log(“MONITORING|1489795335|27.4|latency|user-api-latency”);
console.log(“MONITORING|1489795335|8|count|yubls-served”);
timestamp metric value
metric type
metric namemetrics
logs
CloudWatch Logs AWS Lambda
ELK stacklogs
metrics
CloudWatch
CloudWatch Logs
CloudWatch Logs AWS Lambda
ELK stacklogs
metrics
CloudWatch
memory usedmemory size
billed duration
http://bit.ly/2gGredx
http://bit.ly/2goFZ8F
DASHBOARDS
DASHBOARDS
SET ALARMS
DASHBOARDS
SET ALARMS
TRACK APP-LEVELMETRICS
Not Only CloudWatch
don’t put all your eggs in one basket
aka. you don’t want your monitoring system to fail at the same time as the systems it monitors
CONFIG MANAGEMENT
Lambda
me
Environment variables make it hard to share configurations
across functions.
me
Environment variables make it hard to implement fine-grained
access to sensitive info.
http://bit.ly/2uQKABA
couples ability to deploy with access to sensitive data, which often don’t overlap in a large
engineering team or in a regulated environment
CENTRALISEDCONFIG SERVICE
config servicegoes here
Why not consul or etcd?
! multiple EC2 instances in multi-AZ for HA! have to manage servers, patch OS, patch software, etc.! learning curve for configuring the service! learning curve for using the CLI tools
SSM Parameter
Store
SSM Parameter Store
HTTPS
role-based access
encrypted in-flight
SSM Parameter Store
encrypt
role-based access
SSM Parameter Store
encrypted at-rest
HTTPS
role-based access
SSM Parameter Store
encrypted in-flight
SSM Parameter Store
decrypt
role-based access
CENTRALISEDCONFIG SERVICE
CLIENT LIBRARY
Requirements for client library
! standardise and encapsulate how you manage configs! supports client-side caching (fetch & cache at coldstart)! invalidate cache at interval! invalidate cache explicitly when staleness is detected
http://bit.ly/2yLUjwd
PRO TIPS
max 75 GB total deployment package size*
* limit is per AWS region
Janitor Monkey
disable versionFunctions in
install Serverless framework as dev dependency at project level
dev dependencies are excluded since 1.16.0
http://bit.ly/2vzBqhC
http://amzn.to/2vtUkDU
UNDERSTANDCOLDSTARTS
Amazon X-Ray1st invocation
2nd invocation
cold start
source: http://bit.ly/2oBEbw2
http://bit.ly/2rtCCBz
me
C# and Java experiences ~100 times the cold start time of Python and also suffer from
much higher standard deviation
me
memory size improves cold start time linearly
AVOIDCOLDSTARTS
CloudWatch Event AWS Lambda
CloudWatch Event AWS Lambda
ping
ping
ping
ping
CloudWatch Event AWS Lambda
ping
ping
ping
ping
CloudWatch Event AWS Lambda
ping
ping
ping
ping
HEALTH CHECKS?
AWS Lambda docs
Take advantage of container re-use to improve the performance of your function. Make sure any
externalized configuration or dependencies that your code retrieves are stored and referenced locally after initial execution. Limit the re-initialization of variables/objects on
every invocation. Instead use static initialization/constructor, global/static variables and singletons. Keep alive and reuse connections (HTTP, database, etc.) that
were established during a previous invocation.
http://amzn.to/2jzLmkb
max 5 mins execution time
http://bit.ly/2w6ItdI
CONSIDERPARTIAL
FAILURES
AWS Lambda docs
AWS Lambda polls your stream and invokes your Lambda function.
Therefore, if a Lambda function fails, AWS Lambda attempts to process the erring batch of records until the time
the data expires.
http://amzn.to/2vs2lIg
vsprocessing halts until failed
events are retried successfully/expired from stream
prioritize realtime-ness, retry failed events with best effort,
then skip
SNS
Kinesis
SQS
after 3 attempts
share processing logic
events are processed in chronological order
failed events are retried out of sequence
PROCESS SQSWITH RECURSIVE
FUNCTIONS
http://bit.ly/2npomX6
AVOID HOTKINESS
STREAMS
AWS Lambda docs
Each shard can support up to 5 transactions per second for
reads, up to a maximum total data read rate of 2 MB per second.
http://amzn.to/2ubyaot
AWS Lambda docs
If your stream has 100 active shards, there will be 100 Lambda functions running concurrently. Then, each Lambda function
processes events on a shard in the order that they arrive.
http://amzn.to/2ubyaot
when no. of processors goes up…
ReadProvisionedThroughputExceeded
can have too many Kinesis read operations…
ReadRecords.IteratorAge
unpredictable spikes in read ‘latency’…
can kinda workaround…
http://bit.ly/2uv5LsH
clever, but costly
new tool, new problemsbut they’re easier to deal with
@theburningmonktheburningmonk.comgithub.com/theburningmonk
http://bit.ly/2yQZj1H