When lambdas haven't been hit for 15mins the first hit after has a noticeably longer start time. It's due to deprovisioning/reprovisioning underlying resources like network interfaces. Some people do crazy stuff like a scheduled task to hit their own service to combat this so AWS promised to solve it.
Even if you invoke your lambda function to warm it up in anticipation of traffic, you'll still hit cold starts if the lambda needs to scale out; the new machines are exposed to inputs "cold." Those crazy patterns trying to warm the lambda up are really crazy if you think about it because no one is using them is really aware of the underlying process(es) involved.
"Why are you throwing rocks at that machine?"
"It makes it respond more quickly to actual client requests. Sometimes."
"Sometimes?"
"Well, most the time."
"Why's that? What's causing the initial latency?"
"Cold starts."
"Yeah, but what's that mean?"
"The machine is booting or caching runtime data or, you know, warming up and serverless. Anyway, don't think about it too much, just trust me on this rock thing. Speaking of which, I got to get back to bastardizing our core logic. Eddie had great results with his new exponential backup rock toss routine and I'm thinking of combining that with a graphql random parameter generator library that Ted said just went alpha this afternoon."
In addition to what the sibling reply said. There is also the issue of your choice of runtimes. Java has the slowest startup time and Go the fastest. C# is faster than Java but still slower than Python and Node.
Clever uses of dependency injection without reflection (think dagger not spring) and reducing code size as much as possible can give you fairly fast cold start times with a Java execution environment. Bloated classpaths and reflection filled code will destroy your initialization time.
Or is it when deploying new code?