Skip to content

RDS PostgreSQL 17.11 (and 18.6): ~27ms per-backend process startup vs ~4.5ms on 14.19 — connection-heavy workloads see 3–4x CPU

0

Starting a new backend process costs ~27 ms of CPU on RDS PG 17.11, versus ~4.5 ms on 14.19. Because PostgreSQL forks one process per connection, and our API opens 40–120 new connections/second, that fixed cost is paid continuously and pins the CPU.

Isolation steps, all reproducible:

  • On an idle PG17.11 instance, launching a single parallel worker for SELECT 1 takes 28–585 ms; serial SELECT 1 is fast. On PG14.19 the same parallel worker launch is 3.8–4.7 ms.
  • The idle PG17.11 instance shows a ~27–29 ms floor for any process start vs ~4 ms on PG14.19 — a fixed per-process tax, independent of our data.
  • Reproduced on a brand-new empty 17.11 instance, default parameter group, no pglogical, wal_level=minimal, 81 max_connections: 26.5–29.5 ms per process start — same as prod and dev. So it is not our schema, data, or configuration.

It is not specific to a single major version — it arrived in the August 2026 minors

Same test across PG18 minors on RDS:

versionper-process start
18.44.6–5.0 ms
18.625.9–27.1 ms
14.19~4.5 ms
17.1126–29 ms

The regression appears exactly at the August 2026 minor releases (17.11, 18.6), and 18.4 is clean. That points at a change in the RDS build shared across branches, not at PostgreSQL core code. Our leading suspect is the switch to AWS-LC-FIPS 2025, whose CPU Jitter Entropy source carries a one-time initialization cost per process; since Postgres forks per connection, that cost recurs on every new backend.

Questions for AWS

  1. Is the ~22 ms per-process startup increase on RDS PG 17.11 / 18.6 a known regression, and is it linked to the AWS-LC-FIPS jitter-entropy change in the August 2026 build?
  2. Is a fix expected in a future minor, or is this the new baseline?
  3. Is there a parameter or configuration to seed the RNG once at postmaster start instead of per forked backend?

Mitigations we applied (for others hitting this)

  • Raised the app connection-pool idle timeout so pooled connections stop being recycled every 1 s — cut new-connection rate from ~124/s to ~57/s and CPU by ~a third.

Environment: RDS for PostgreSQL, eu-central-1, db.m5.2xlarge, upgraded via Blue/Green 14.19 → 17.11 on 2026-09-18.

2 Answers
1

Hello.

Based on your test results, it is highly likely that the AWS-LC-FIPS 2025 CPU Jitter Entropy initialization is involved.
There is an existing performance-related issue in AWS-LC similar to this phenomenon.
Although it has already been closed, it is possible that performance degradation could still occur.
https://github.com/aws/aws-lc-rs/issues/899

The aws-lc-rs README notes a "one-time-per-process" latency cost associated with CPU Jitter Entropy—typically around 50 ms—so it is certainly possible that aws-lc-rs is the cause of the slowdown.
https://github.com/aws/aws-lc-rs/blob/main/book/src/resources.md

When set to 1, disables the CPU jitter entropy source in the build. This affects the random number generation subsystem. Use of jitter entropy has a one-time-per-process latency cost, typically around 50ms, for the collection of entropy. This flag may be used to eliminate this latency.

EXPERT

answered 2 months ago

  • Hello. Thanks for sharing. I'm pretty sure that tha's the cause I hope AWS will fix it for RDS

0

The ~27 ms per-backend process startup cost you're observing on RDS PostgreSQL 17.11 and 18.6 (versus ~4.5 ms on 14.19) is indeed significant for connection-heavy workloads. Each PostgreSQL backend is a full process that requires forking, copying memory structures, initializing process state, and authentication. When your API opens 40–120 new connections per second, this fixed startup cost is paid continuously and can lead to substantial CPU consumption.

Your observation that the regression appeared specifically in the August 2026 minor releases (17.11 and 18.6) while 18.4 remained clean is valuable diagnostic information. The timing and cross-version nature of the issue does suggest an infrastructure-layer change rather than a PostgreSQL core regression.

Regarding process forking overhead, parallel query execution provides a useful analogy for understanding the costs involved. Each parallel worker is a full PostgreSQL backend process requiring process forking and authentication, consuming connection slots from your max_connections limit. Each worker also consumes its own memory, including work_mem for operations. Under high concurrency, this can lead to CPU saturation, increased context switching, and connection exhaustion.

For your specific situation with connection-heavy workloads, your mitigation strategy of raising the connection-pool idle timeout to reduce the new-connection rate from ~124/s to ~57/s is sound. Additional strategies to consider include implementing connection pooling if not already in place, as frequent opening and closing of database connections without pooling can cause elevated authentication attempts and slow connection establishment times. You might also evaluate whether your application can maintain longer-lived connections or use a connection pooler like PgBouncer or RDS Proxy to reduce the frequency of new backend process creation.

For your specific questions about the AWS-LC-FIPS change and whether this represents a new baseline, I recommend opening a support case through the AWS Support Center at https://console.aws.amazon.com/support/ where the RDS team can investigate the specific build changes in those minor releases and provide guidance on any available mitigations or upcoming fixes. When you add correspondence to your support case, include the detailed version comparison data and timing information you've gathered, as this will help the team investigate efficiently.
Sources
Initial troubleshooting for common PostgreSQL performance issues in Aurora PostgreSQL - Amazon Aurora
Best Practices for Parallel Queries in RDS for PostgreSQL - Amazon Relational Database Service

answered 2 months ago

AWS
SUPPORT ENGINEER

reviewed 2 months ago

EXPERT

reviewed 2 months ago

You are not logged in. Log in to post an answer.

A good answer clearly answers the question and provides constructive feedback and encourages professional growth in the question asker.