perf(lambda-at-edge): fix rollup bundling for dynamic imports perf

[dphang]

Note: testing rollup config to properly bundle dynamic imports like s3client which is not always needed. The cost is an additional wrapper on top of main code but this seems expected optimization from Rollup.js (preloading and parsing optimization).

Previously, inlining dynamic imports into one file was still causing that code to be at least partially preloaded outside of the handler, e.g initializing some S3/AWS SDK related objects even if the S3Client itself was only initialized when needed. So it caused higher cold starts due to this for all paths. With this change, we split into multiple chunks. Total there should be a few chunks:

• Entry chunk (wrapper of ~20 lines on top of other code)
• Core chunk (all of the routing code bundled into one file)
• Dynamic chunks (e.g S3 client, GetObjectCommand, PutObjectCommand) which is heavy code only needed on fallback paths

Total size of all chunks remains around the same but slightly higher at 657 kB vs. 651 kB before. However, Lambda does not need to read (and preload some parts of it) ~166 kB of code for S3 unless it's needed. There is one extra file I/O due to the wrapper requiring the core code but from tests, it appears to be negligible as ~25 ms is saved in init time from my tests (195 -> 170 ms with the 512 MB Lambda function)
In the future, we can do further optimizations like getting rid of complexity of having the origin response handler, and even using Rollup again at build/deploy time to further get rid of code paths that are not used (e.g if user has no rewrites, we can get rid of that code)

[codecov]

Codecov Report

Merging #1029 (ebbb63a) into master (c86bd09) will not change coverage.
The diff coverage is 100.00%.

❗ Current head ebbb63a differs from pull request most recent head 244853c. Consider uploading reports for the commit 244853c to get more accurate results

@@           Coverage Diff           @@
##           master    #1029   +/-   ##
=======================================
  Coverage   82.31%   82.31%           
=======================================
  Files          69       69           
  Lines        2578     2578           
  Branches      615      615           
=======================================
  Hits         2122     2122           
  Misses        390      390           
  Partials       66       66           

Impacted Files	Coverage Δ
packages/libs/lambda-at-edge/src/build.ts	95.33% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update c86bd09...244853c. Read the comment docs.

[jvarho]

Yes, this looks better than my effort 👍

I can confirm this reproduces the improvement I saw in cold starts. Here are five-run averages for before and after this branch. I'm using 128 MB memory setting and a redirect to get the worst case.

Before:

Max Memory Used: 71 MB
Init Duration: 201.8 +/- 8.3 ms
Duration: 30.2 +/- 5.3 ms

After:

Max Memory Used: 67 MB
Init Duration: 161.2 +/- 6.9 ms
Duration: 31.3 +/- 6.7 ms

About 20% lower init duration and a bit less memory use.

I have not tested the fallback time, because it's dominated by S3 latency and rendering.