Motivation

While deploying our re:Invent demo stack I found it takes ~85 seconds for a warm deploy on my system. This time is dominated by deploying lambda functions which each take 5 seconds.

When we deploy resources, we create a loop where

• the first iteration deploys the resource
• the remaining iterations poll the resource for completion

AWS CloudFormation waits between polling loops (to reduce burden on their networking and services). On LocalStack we don't need to do this - a lot of our resources are in-memory so are instantly available. Therefore we set the sleep time for the first iteration of this loop only to 0 seconds, and wait 5 seconds for every other iteration.

Since the lambda functions take a minimum of 5 seconds each, I suspect the deployment loop is going round more than twice.

Increasing the number of loop iterations that we don't sleep for would reduce this deployment time greatly.

Changes

• Add a configurable CFN_NO_WAIT_ITERATIONS to control the number of loop iterations we don't wait 5 seconds between, defaulting to 5 iterations
• This change was added to the legacy provider since the resource provider executor is shared between the two
• Decrease the default sleep time from 5 to 1 second, which should speed everything up (thanks @dfangl for the suggestion)
• Now that we are iterating quicker, I have increased the minimum number of iterations to 10 (if for some reason the user specifies a CFN_PER_RESOURCE_TIMEOUT of e.g. 0!)

After these changes, our re:Invent demo stack takes ~30 seconds, leading to a 2.8x total speedup 🎉

Tests

• Test deploying a large stack with lambda functions with this setting at its previous default of 1, and with the new default of 5 and observe the total time to deploy has decreased

Related

Resolves UNC-122