Storage

The filesystem on AWS Lambda is read-only, except for the /tmp directory.

On top of that, the filesystem is not shared between instances of a lambda when it scales up. For example a file /tmp/foo.json will not be shared across instances of the same lambda. Since a lambda can scale up or down at any time, data in the /tmp directory can be lost.

Application data

Application data must not be stored in /tmp because of the behavior described above.

Instead, data can be stored in databases or in storage services like AWS S3.

S3 storage

It is possible to deploy a S3 bucket in serverless.yml using the Storage feature of the Lift plugin. For example:

# serverless.yml

# ...

provider:
    environment:
        BUCKET_NAME: ${construct:reports-bucket.bucketName}

constructs:
    reports-bucket:
        type: storage

Read more in the Lift documentation.

Application cache

Performance-wise, using AWS S3 for storing the application cache is not ideal.

The following solutions can be used instead:

  • pre-generate the cache in the project directory before deploying
  • store the cache into the /tmp directory
  • store the cache into a distributed cache service like Memcache, Redis or DynamoDB (an AWS service which fits the pay-per-request model nicely)

Pre-generating the cache

Some frameworks allow to pre-generate some caches. For example in Symfony the container can be compiled via bin/console cache:warmup, or in Laravel the config cache can be generated before deploying.

When possible this is the best solution: no generation will occur in production, and reading from the filesystem will be fast.

Store in the /tmp directory

Some framework or library caches must be written into files. In that case storing in the /tmp directory is a good solution.

Remember that anything stored in /tmp will be lost when a lambda stops. When a lambda starts, the /tmp directory will be empty so the cache will be generated again.

Note that this is useful for deployments: no need to clear caches on deployments since a new version of the lambda will run on new instances (with an empty /tmp directory).

This solution is ideal when the cached data is fast to generate and never changes (e.g. template caching, framework caches).

Store in a distributed cache service

Using a distributed cache service has the following advantages:

  • the cache is not lost when the lambda scales down
  • the cache is not lost when deploying
  • the cache is shared between all lambda instances

The disadvantage is that if the data format changes between deployments, then a deployment strategy must be used to either clear the cache and regenerate it, or separate the cache between application versions.

Cache services that can be used include for example Redis, Memcache or DynamoDB. AWS offers those as managed services through AWS ElastiCache (Redis, Memcache) or DynamoDB.

Note that Redis and Memcache (through ElastiCache) run even when not used, which incurs costs. DynamoDB is a little slower than both of those but can be deployed in a "pay-per-request" mode where costs are proportional to the usage. There is a package implementing the PSR cache interfaces using DynamoDB (rikudou/psr6-dynamo-db).

This solution is ideal for cache data that can change during the life of the application (e.g. caching a website menu, an API response…).

Deploying DynamoDB tables

As any service, DynamoDB tables can be deployed via CloudFormation using the resources key in serverless.yml:

service: app
...

resources:
    Resources:
        CacheTable:
            Type: AWS::DynamoDB::Table
            Properties:
                AttributeDefinitions: # only keys are defined here, other attributes are dynamic
                    -   AttributeName: id # adds a mandatory id field
                        AttributeType: S # the type of id is a string
                BillingMode: PAY_PER_REQUEST # billed for each request instead of paying for a constant capacity
                TimeToLiveSpecification: # deletes cache keys automatically based on a ttl field which contains a timestamp
                    AttributeName: ttl
                    Enabled: true
                KeySchema:
                    -   AttributeName: id
                        KeyType: HASH # the type of key, HASH means partition key (similar to primary keys in SQL)

We need to allow code in Lambda functions to access DynamoDB. We can also pass the table name as an environment variable to the application.

service: app
provider:
    iam:
        role:
            statements:
                - Effect: Allow
                    Resource: !GetAtt CacheTable.Arn
                    Action:
                        - dynamodb:DescribeTable
                        - dynamodb:Query
                        - dynamodb:Scan
                        - dynamodb:GetItem
                        - dynamodb:PutItem
                        - dynamodb:UpdateItem
                        - dynamodb:DeleteItem
    environment:
        # That environment variable will contain the table name
        DYNAMODB_TABLE: !Ref CacheTable