pageserver: coalesce read paths #7477

VladLazar · 2024-04-23T11:49:02Z

Problem

We are currently supporting two read paths. No bueno.

Summary of changes

High level: use vectored read path to serve get page requests - gated by get_impl config
Low level:

Add ps config, get_impl to specify which read path to use when serving get page requests
Fix base cached image handling for the vectored read path. This was subtly broken: previously we
would not mark keys that went past their cached lsn as complete. This is a self standing change which
could be its own PR, but I've included it here because writing separate tests for it is tricky.
Fork get page to use either the legacy or vectored implementation
Validate the use of vectored read path when serving get page requests against the legacy implementation.
Controlled by validate_vectored_get ps config.
Use the vectored read path to serve get page requests in tests (with validation).

Note

Since the vectored read path does not go through the page cache to read buffers, this change also
amounts to a removal of the buffer page cache. Materialized page cache is still used.

Deployment Plan

Make sure this is correct.
1.1. Run test suite with validation - DONE
1.2. Run a prodlike bench with validation in staging
Understand perf degradation (if any)
2.1. Run checked in benchmarks - DONE here - looks ok, didn't notice significant degradations
2.2. I'll manually run test_pageserver_max_throughput_getpage_at_latest_lsn - DONE here - looks better than the current read path
2.2. Run prodlike bench without validation in staging
Staged rollout in prod without validation

Related #7381

Checklist before requesting a review

I have performed a self-review of my code.
If it is a core feature, I have added thorough tests.
Do we need to implement analytics? if so did you add the relevant metrics to the dashboard?
If this PR requires public announcement, mark it with /release-notes label and add several sentences in this section.

Checklist before merging

Do not forget to reformat commit message to not include the above checklist

github-actions · 2024-04-23T11:59:32Z

2766 tests run: 2645 passed, 0 failed, 121 skipped (full report)

Code coverage* (full report)

functions: 28.1% (6486 of 23064 functions)
lines: 47.0% (46102 of 98113 lines)

* collected from Rust tests only

_{The comment gets automatically updated with the latest test results
fed8579 at 2024-04-24T15:30:19.671Z :recycle:}

skyzh

LGTM

VladLazar · 2024-04-25T11:07:37Z

I reran test_pageserver_max_throughput_getpage_at_latest_lsn to compare the implementations:
Interestingly when vectored get is enabled, the benchmark achieves higher throughput and lower mean latencies, but higher tail latencies. I think this can be explained by the fact that we are not using the buffer cache in the vectored implementation. What's surprising is that the vectored implementation performs well even with a cold linux page cache.

Non vectored get - baseline

[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 1
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 6
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 114762
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.260 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.416 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 0.450 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 0.536 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 1.018 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 1
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 13
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 108648
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.275 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.376 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 0.409 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 0.510 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 0.979 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 10
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 6
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 425238
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.704 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 1.162 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 1.404 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 1.869 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 3.939 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 10
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 13
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 473836
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.632 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.939 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 1.128 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 1.542 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 2.875 ms

Vectored get - cold page page cache

[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 1
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 13
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 136567
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.218 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.516 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 0.747 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 1.307 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 2.443 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 10
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 6
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 504282
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.594 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 1.042 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 1.544 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 2.751 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 5.391 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 10
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 13
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 498062
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.601 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 1.005 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 1.394 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 2.271 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 4.207 ms

Vectored get - hot page cache

[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 1
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 6
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 360067
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.082 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.130 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 0.425 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 0.492 ms
[release-pg15-1-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 0.921 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 1
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 13
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 156694
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.190 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.384 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 0.421 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 0.479 ms
[release-pg15-1-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 0.911 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 10
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 6
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 547575
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.547 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.897 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 1.101 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 1.605 ms
[release-pg15-10-6-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 5.243 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.n_tenants: 10
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pgbench_scale: 13
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.duration: 30 s
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.page_cache_size: 134217728 byte
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.pageserver_config_override.max_file_descriptors: 500000
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.request_count: 575608
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_mean: 0.520 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p95: 0.782 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99: 0.966 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.9: 1.458 ms
[release-pg15-10-13-30].pageserver_max_throughput_getpage_at_latest_lsn.latency_percentiles.p99.99: 2.963 ms

VladLazar · 2024-04-25T11:08:11Z

At this point I think we should merge and try it out in staging.

VladLazar force-pushed the vlad/coalesce-read-paths-test branch from d52eb01 to 0edc368 Compare April 23, 2024 11:51

VladLazar force-pushed the vlad/coalesce-read-paths-test branch 3 times, most recently from 064d1c8 to 1826b13 Compare April 23, 2024 17:45

VladLazar changed the title ~~[DNM] coalescing read paths testing~~ pageserver: coalesce read paths Apr 23, 2024

VladLazar mentioned this pull request Apr 23, 2024

Coalesce non-vectored get with vectored code (use vectored get for single key gets) #7381

Closed

VladLazar added run-benchmarks Indicates to the CI that benchmarks should be run for PR marked with this label and removed run-benchmarks Indicates to the CI that benchmarks should be run for PR marked with this label labels Apr 24, 2024

VladLazar added 9 commits April 24, 2024 15:40

pageserver/config: add config for singular get impl

d9d37ab

pageserver: fix vectored read path cache handling

049dcb4

pageserver: use vectored get for singular gets when configured

a8a1e47

pageserver: validate singular vectored get results

35885f9

pageserver: improve error parity between read paths

c608159

test: use vectored impl for singular gets in tests

e8a32d7

test: skip validation for debug mode in some demanding tests

9d64439

tests: unset new config for forward compat test

478d999

tests: handle not found errors from both read paths in lsn mapping test

fed8579

VladLazar force-pushed the vlad/coalesce-read-paths-test branch from ff94f16 to fed8579 Compare April 24, 2024 14:44

VladLazar marked this pull request as ready for review April 24, 2024 17:04

VladLazar requested a review from a team as a code owner April 24, 2024 17:04

VladLazar requested review from skyzh and jcsp April 24, 2024 17:04

skyzh approved these changes Apr 24, 2024

View reviewed changes

VladLazar merged commit e4a279d into main Apr 25, 2024
57 checks passed

VladLazar deleted the vlad/coalesce-read-paths-test branch April 25, 2024 12:29

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

pageserver: coalesce read paths #7477

pageserver: coalesce read paths #7477

VladLazar commented Apr 23, 2024 •

edited

Loading

github-actions bot commented Apr 23, 2024 •

edited

Loading

skyzh left a comment

VladLazar commented Apr 25, 2024

VladLazar commented Apr 25, 2024

pageserver: coalesce read paths #7477

pageserver: coalesce read paths #7477

Conversation

VladLazar commented Apr 23, 2024 • edited Loading

Problem

Summary of changes

Note

Deployment Plan

Checklist before requesting a review

Checklist before merging

github-actions bot commented Apr 23, 2024 • edited Loading

2766 tests run: 2645 passed, 0 failed, 121 skipped (full report)

Code coverage* (full report)

skyzh left a comment

Choose a reason for hiding this comment

VladLazar commented Apr 25, 2024

Non vectored get - baseline

Vectored get - cold page page cache

Vectored get - hot page cache

VladLazar commented Apr 25, 2024

VladLazar commented Apr 23, 2024 •

edited

Loading

github-actions bot commented Apr 23, 2024 •

edited

Loading