AI Answer | Zyeon UI

ready · sources row, cited prose, follow-ups — cross-highlight both ways

Why did our p99 latency triple after moving the API to Lambda?

Answer grounded in 5 sources

Sources· 5

docs.aws.amazon.com1Lambda execution environment lifecycle — init, invoke, shutdown
docs.aws.amazon.com2Concurrency scaling behaviour: burst limits and the request queue
crunchydata.com3Connection pooling when every request might be a new process
docs.aws.amazon.com4RDS Proxy: pinning, multiplexing, and what it actually saves you
news.ycombinator.com5Ask HN: what did serverless actually do to your tail latency?

Three effects usually stack here, and only one of them is the cold start everybody blames. A Node cold start costs roughly 200–400 ms before your handler's first line runs, and it is paid once per execution environment rather than once per request — which is why it shows up in the tail and not in the median.

What actually moves p99

Connection setup: every new environment opens its own database connection, so a burst of 200 concurrent requests becomes 200 handshakes against a pool sized for 20.
Scaling gaps: traffic arriving faster than the burst allowance waits while environments are created, and that wait lands entirely in the tail.
Network position: responses now leave from outside your VPC, so anything large pays a different egress path than the old service did.

The change that moves the number most is usually a connection proxy in front of the database, not more provisioned concurrency. Teams that put RDS Proxy or a pgBouncer-style pooler in the path report p99 falling back to within about 15% of the pre-migration baseline, with provisioned concurrency shaving the remainder.

Keep going

Hover a numbered chip to light up its source card — and a source card to light up the sentences it backs. Click a follow-up and it arrives here.

streaming · streaming — sources landed first, the prose stops mid-sentence behind a caret

Why did our p99 latency triple after moving the API to Lambda?

Writing the answer…

Sources· 5

Three effects usually stack here, and only one of them is the cold start everybody blames. A Node cold start costs roughly 200–400 ms before your handler's first line runs, and it is paid once per execution environment rather than once per request — which is why it shows up in the tail and not in the median.

What actually moves p99

Connection setup: every new environment opens its own database connection, so a burst of 200 concurrent requests becomes 200 handshakes against a pool sized for 20.
Scaling gaps: traffic arriving faster than the burst allowance waits while environments are

empty · empty — documents were retrieved, none of them answered the question

What is the internal SKU code for our 2027 enterprise tier?

No answer for this question.

Sources· 2

Nothing in the retrieved material answers this — rather than guess, the model said nothing.

Keep going

The narrower question is the way out of an empty answer.

loading · loading — the question is already on screen, everything else is a ghost

Why did our p99 latency triple after moving the API to Lambda?

Searching…

error · error — the run failed; Try again switches this card to the ready answer

Why did our p99 latency triple after moving the API to Lambda?

The answer could not be produced.

Couldn't answer that

The model stopped before it produced an answer.

edge · edge — a query that wraps twice, and a citation to a source retrieval never returned

After we migrated the checkout API from the long-lived Fargate service to per-request Lambda functions behind API Gateway, why did the 99th percentile latency triple while the median stayed flat?

Answer grounded in 3 sources

Sources· 3

The median is served by warm environments that already hold a database connection, so it barely moves. The tail is served by environments that do not exist yet, and an internal benchmark that is not in the retrieved set is cited here anyway.9 (source unavailable)