At ipdata.co we use the same APIG+Lambda setup replicated in 11 regions to have the lowest latencies globally. We had to do some extra work to get API keys and rate limiting working but it was worth it. Our setup averages ~44ms response times -
https://status.ipdata.co/.
We wrote about our setup in detail on the Highscalability blog [1]
A few things have changed since we wrote that article;
- We implement custom authorizers, which have helped lower our costs and the auth caching means authentication only happens once per x minutes and all subsequent requests are much faster.
- We use redis and a couple of Kinesis consumers running on a real server to sync data across all our regions. This setup has been battle tested and has successfully processed more than a hundred million API calls in a single day in near real time. [Use pipes and mget in redis for speed]
Here are some answers to a few specific things you raise in your question;
1. Use Sentry for lambda for error handling. The logs you get are incredibly detailed and have single handedly given us the greatest visibility into our application, moreso than any other tool we've tried (like AWS Xray).
2. Cloudwatch logs are tough. You might want to consider piping your logs to an Elasticsearch cluster, that might be a bit costly if you use AWS's Elasticache.
3. We use terraform for deploying our lambda functions and other resources. I'd strongly recommend it.
[1] https://highscalability.com/blog/2018/4/2/how-ipdata-serves-...