lambda duration spike when new cloudwatch log stream

Question

The problem

I have an aws lambda that runs by ~2ms of runtime duration. Each time a new cloudwatch log stream in that lambda's log-group is created, the first run of that lambda takes ~15K ms (x7500 factor).

I'm looking for a way to make the lambda permanently run in ~2ms, and avoid those duration spikes completely.

Investigation details

1. I thought that it is related to lambda warmup on behind-the-scenes container allocation, so I added a event-bridge trigger by constant rate of every 3 minutes to that lambda. Also, I allocated provisioned-concurrency for that lambda. But it didn't solve the issue.
2. I have another lambda, both are using the same code but running in a different flow and it doesn't happen to it.
3. Note that currently there is no external traffic, only the event-bridge triggering which has a controlled rate.
4. If it's relevant: I'm using python3.7 with NLTK lib in both lambdas but the delayed one is using wordnet's lemmatization (statically defined as recommended here).