lambda-mutex

A mutex implementation for [AWS Lambda] (https://aws.amazon.com/lambda/), leveraging AWS DynamoDB

Usage

Let's say you want to ensure that only 1 lambda function can access a resource (for example an instance i-8abd82c31) at a time

   from dyndbmutex.dyndbmutex import DynamoDbMutex
   # at the beginning of your lambda handler
   # generate a unique name for this instantiation of lambda
   my_name = str(uuid.uuid4()).split("-")[0]
   m = DynamoDbMutex(lockname='i-8abd832c32', holder=my_name, timeoutms=20 * 1000)
   locked = m.lock()
   if locked:
      # critical section begin
       ......
      # critical section end
      m.release()

You can also use the with pattern:

   from dyndbmutex.dyndbmutex import DynamoDbMutex, AcquireLockFailedError
   my_name = str(uuid.uuid4()).split("-")[0]
   m = DynamoDbMutex('i-8abd832c32', my_name, 20 * 1000)
   try:
       with m:
          # critical section
   except mutex.AcquireLockFailedError:
       #m will be released at this point

Theory of operation

Uses [DynamoDb conditional write] (http://docs.aws.amazon.com/amazondynamodb/latest/developerguide/WorkingWithItems.html#WorkingWithItems.ConditionalUpdate) as a atomic compare-and-swap operation to implement a mutex.

• Prune lock: if the acquirer fails to release it within the timeout, release it if it is expired
• Acquire lock: prune the lock if required. If the lock is now released, acquire it
• Release lock: release it if I am the holder, otherwise fail.

Since the conditional write is atomic (test and set), this works very well. In fact the code doesn't even read the table, only writes to it. (We could even make the lock re-entrant since we have the owner/holder information, but this is not likely to be needed in the lambda usage pattern)

Setup

The default name for the Mutex table in DynamoDB is 'Mutex'. You can change this by setting an environment variable:

export DD_MUTEX_TABLE_NAME=FancyPantsMutex

The code will auto-create the mutex DynamoDB table, but this could take at least 20 seconds. Use the create-table script in the scripts directory before using this mutex library.

Notes and Limitations

Although the code is general-purpose and can be used outside of AWS lambda, note the following limitations:

• Not designed for fine-grained parallelism. Generally, it is expected that you acquire a lock and hold it for the duration of the lambda function
• Does not detect/prevent deadlocks. There is no spin lock, but the mutex user could create one by spinning until an acquire succeeds. Generally, one should avoid taking more than 1 lock in a lambda.
• Not re-entrant. If a thread (lambda function) tries to re-acquire a lock it already holds, it will block
• Not designed for speed. The DynamoDb table backing the locks is generally provisioned as low throughput (2 ops/sec)
• No cleanup. Each named mutex gets a row in the Dynamodb table. These rows are never cleaned up even if nobody is using the mutex. It should be trivial to write a lambda to vaccuum these rows, however, given the small size of the row and the infinite capacity of DynamoDb, it is not likely to be a problem, if used for its intended purpose.

TODO

• No limits on timeout. Perhaps there should be one (300 seconds?)