It never fails. Every time I talk about serverless, someone pushes back with the cold start argument. I still see it in forums, in blog comments, i...
For further actions, you may consider blocking this person and/or reporting abuse
Cold starts being 'dead' depends heavily on your runtime. Java on Lambda still has a real conversation to have here — even with SnapStart, the restore latency on complex Spring Boot contexts isn't zero. The story for Node or Python runtimes? Yeah, pretty much solved. But if you're running anything JVM-based with a non-trivial dependency graph, I'd be more cautious about declaring victory. Curious if you benchmarked anything beyond the usual Node/Python suspects.
It’s wild to see how much progress has been made with things like tiered compilation and proactive initialization to the point where "cold starts" aren't the boogeyman they used to be. It definitely shifts the conversation more toward optimizing actual code logic rather than just fighting the platform.
This is compelling
Cold starts still exist, but they’re not the primary bottleneck anymore for most workloads. Architecture decisions probably matter more now than runtime latency.
That's actually good to know!
Great post!
The Rust numbers are wild. 14ms on arm64 is basically
free.
I picked Rust for a blockchain node project and the
performance predictability has been the biggest win.
Not just raw speed but the fact that you don't get
random pauses. No GC, no JIT warmup, no hidden
allocation spikes. When you need deterministic block
times in a consensus engine that's the thing that
actually matters.
The arm64 improvement across every runtime is
interesting too. 17-25% faster for just switching
architecture with zero code changes is hard to argue
against. Curious if you've seen similar gains on
Graviton for longer running workloads or if it's
mostly a cold start benefit.
Wow this is really changing my mental model about Lambda, thanks Eric!
How are 150ms coldstarts acceptable? This is insane lol, cold starts are still very much alive :D
Ha! That's kind of the whole point of the post though. If 150ms is "insane" to you, what were you dealing with in 2018 when VPC-connected Lambda added 10-15 seconds?
150ms is less than a single frame at 60fps. Your users can't perceive it. Your monitoring dashboards won't flag it. It happens on less than 1% of invocations.
Is it zero? No. Is it "alive" in the way it was when people had legitimate reasons to avoid serverless? Also no. That's the argument — not that cold starts literally don't exist, but that they've moved from "architecture-breaking problem" to "rounding error for most workloads."
If you're building something where 150ms genuinely matters (HFT, real-time bidding, competitive gaming), the post covers that too. For a web app, an API, a chat service? You won't notice.
Sorry, but respectfully disagree.
Thats not the whole story? You still need to do your actual processing + latency of the user. The difference between 300ms and 450ms is VERY noticeable to some audiences (especially if you sell to developers like I do). I think 150ms matter in a lot more usecases than HFT. Obviously biased (I am the Founder of a company that does the exact opposite to serverless!).
Also, I think its a bit ironic that Amazon was the company that popularized the whole "100ms costs you 1% in revenue" story before Lambda was a thing :D
Fair enough and I appreciate it!
Thank you