fix: further optimize archive workflow listing. Fixes #13601 #13819

MasonM · 2024-10-26T01:38:02Z

Motivation

Listing archived workflows can be slow if you have a very large number of workflows (100,000+), or the average workflow size is high (100KB), even after the optimizations from #13566 and #13779. This makes some additional optimizations that speed up the queries by ~90% on MySQL and ~50% on PostgreSQL.

Modifications

The bottleneck for these queries depended on whether you use MySQL or PostgreSQL, each of which required a different fix. For PostgreSQL, the bottleneck was detoasting overhead, as explained in #13601 (comment). The fix was to use a common table expression to reduce the amount of times workflow needs to be detoasted, as suggested by @kodieg in #13601 (comment). The new query looks like this:

WITH workflows AS (
  SELECT
    "name",
    "namespace",
    "uid",
    "phase",
    "startedat",
    "finishedat",
    coalesce(workflow->'metadata', '{}') as metadata,
    coalesce(workflow->'status', '{}') as status,
    workflow->'spec'->>'suspend' as suspend
  FROM "argo_archived_workflows"
  WHERE (("clustername" = $1 AND "namespace" = $2 AND "instanceid" = $3))
  ORDER BY "startedat" DESC
  LIMIT 100
) (
  SELECT
    "name",
    "namespace",
    "uid",
    "phase",
    "startedat",
    "finishedat",
    coalesce(metadata->>'labels', '{}') as labels,
    coalesce(metadata->>'annotations', '{}') as annotations,
    coalesce(status->>'progress', '') as progress,
    coalesce(metadata->>'creationTimestamp', '') as creationtimestamp,
    "suspend",
    coalesce(status->>'message', '') as message,
    coalesce(status->>'estimatedDuration', '0') as estimatedduration,
    coalesce(status->>'resourcesDuration', '{}') as resourcesduration
  FROM "workflows"
)

For MySQL, the bottleneck was the optimizer inexplicably refusing to use the argo_archived_workflows_i4 index and instead using the primary key, which is much more expensive. As explained by @Danny5487401 in #13563 (comment), two ways of solving that are using FORCE INDEX or adding a union index on (clustername, startedat). Using FORCE INDEX is slightly hacky, and adding a new index is needlessly wasteful when we already have argo_archived_workflows_i4, so I opted to modify that index to cover (clustername, startedat). The new query looks like this:

SELECT
  `name`,
  `namespace`,
  `uid`,
  `phase`,
  `startedat`,
  `finishedat`,
  coalesce(workflow->'$.metadata.labels', '{}') as labels,
  coalesce(workflow->'$.metadata.annotations', '{}') as annotations,
  coalesce(workflow->>'$.status.progress', '') as progress,
  coalesce(workflow->>'$.metadata.creationTimestamp', '') as creationtimestamp,
  workflow->>'$.spec.suspend',
  coalesce(workflow->>'$.status.message', '') as message,
  coalesce(workflow->>'$.status.estimatedDuration', '0') as estimatedduration,
  coalesce(workflow->'$.status.resourcesDuration', '{}') as resourcesduration
FROM `argo_archived_workflows`
WHERE ((`clustername` = ?  AND `namespace` = ? AND `instanceid` = ?))
ORDER BY `startedat` DESC
LIMIT 100

Verification

First, I used #13715 to generate 100,000 randomized workflows, with https://gist.github.com/MasonM/52932ff6644c3c0ccea9e847780bfd90 as a template:

PostgreSQL - time go run ./hack/db fake-archived-workflows --template "@very-large-workflow.yaml" --rows 100000
MySQL - time go run ./hack/db fake-archived-workflows --template "@very-large-workflow.yaml" --rows 100000 -d 'mysql:password@tcp/argo'

Then, I ran make BenchmarkWorkflowArchive once on the main branch and once on this branch (with migration applied), and used benchstat to compare:

PostgreSQL

$ benchstat postgres_before.txt postgres_after2.txt 
goos: linux
goarch: amd64
pkg: github.com/argoproj/argo-workflows/v3/test/e2e
cpu: 12th Gen Intel(R) Core(TM) i5-12400
                                                     │ postgres_before.txt │        postgres_after2.txt         │
                                                     │       sec/op        │   sec/op     vs base               │
WorkflowArchive/ListWorkflows-12                              25.110m ± 1%   9.694m ± 3%  -61.39% (p=0.002 n=6)
WorkflowArchive/ListWorkflows_with_label_selector-12           26.14m ± 3%   11.06m ± 3%  -57.70% (p=0.002 n=6)
WorkflowArchive/CountWorkflows-12                              11.84m ± 1%   11.89m ± 2%        ~ (p=0.310 n=6)
geomean                                                        19.81m        10.84m       -45.26%

                                                     │ postgres_before.txt │        postgres_after2.txt         │
                                                     │        B/op         │     B/op      vs base              │
WorkflowArchive/ListWorkflows-12                              497.2Ki ± 0%   499.2Ki ± 0%  +0.40% (p=0.002 n=6)
WorkflowArchive/ListWorkflows_with_label_selector-12          504.0Ki ± 0%   504.9Ki ± 0%  +0.18% (p=0.002 n=6)
WorkflowArchive/CountWorkflows-12                             8.938Ki ± 1%   8.948Ki ± 2%       ~ (p=0.623 n=6)
geomean                                                       130.8Ki        131.1Ki       +0.23%

                                                     │ postgres_before.txt │         postgres_after2.txt         │
                                                     │      allocs/op      │  allocs/op   vs base                │
WorkflowArchive/ListWorkflows-12                               8.370k ± 0%   8.393k ± 0%  +0.27% (p=0.002 n=6)
WorkflowArchive/ListWorkflows_with_label_selector-12           8.405k ± 0%   8.428k ± 0%  +0.26% (p=0.002 n=6)
WorkflowArchive/CountWorkflows-12                               212.0 ± 0%    212.0 ± 0%       ~ (p=1.000 n=6) ¹
geomean                                                        2.462k        2.466k       +0.18%
¹ all samples are equal

MySQL

$ benchstat mysql_before.txt mysql_after.txt 
goos: linux
goarch: amd64
pkg: github.com/argoproj/argo-workflows/v3/test/e2e
cpu: 12th Gen Intel(R) Core(TM) i5-12400
                                                     │ mysql_before.txt │           mysql_after.txt           │
                                                     │      sec/op      │    sec/op     vs base               │
WorkflowArchive/ListWorkflows-12                           43.510m ± 4%   1.650m ± 20%  -96.21% (p=0.002 n=6)
WorkflowArchive/ListWorkflows_with_label_selector-12       69.646m ± 2%   4.513m ± 24%  -93.52% (p=0.002 n=6)
WorkflowArchive/CountWorkflows-12                           27.37m ± 4%   28.98m ± 24%        ~ (p=0.394 n=6)
geomean                                                     43.61m        5.998m        -86.25%

                                                     │ mysql_before.txt │          mysql_after.txt           │
                                                     │       B/op       │     B/op      vs base              │
WorkflowArchive/ListWorkflows-12                           477.3Ki ± 0%   457.6Ki ± 0%  -4.13% (p=0.002 n=6)
WorkflowArchive/ListWorkflows_with_label_selector-12       488.1Ki ± 0%   461.0Ki ± 0%  -5.54% (p=0.002 n=6)
WorkflowArchive/CountWorkflows-12                          7.937Ki ± 1%   7.944Ki ± 0%       ~ (p=0.461 n=6)
geomean                                                    122.7Ki        118.8Ki       -3.22%

                                                     │ mysql_before.txt │           mysql_after.txt           │
                                                     │    allocs/op     │  allocs/op   vs base                │
WorkflowArchive/ListWorkflows-12                            8.224k ± 0%   7.705k ± 0%  -6.31% (p=0.002 n=6)
WorkflowArchive/ListWorkflows_with_label_selector-12        8.277k ± 0%   7.732k ± 0%  -6.58% (p=0.002 n=6)
WorkflowArchive/CountWorkflows-12                            205.0 ± 0%    205.0 ± 0%       ~ (p=1.000 n=6) ¹
geomean                                                     2.408k        2.303k       -4.35%
¹ all samples are equal

Signed-off-by: Mason Malone <[email protected]>

MasonM · 2024-10-26T01:58:02Z

Makefile

@@ -39,6 +39,7 @@ E2E_PARALLEL          ?= 20
 E2E_SUITE_TIMEOUT     ?= 15m
 GOTEST                ?= go test -v -p 20
 ALL_BUILD_TAGS        ?= api,cli,cron,executor,examples,corefunctional,functional,plugins
+BENCHMARK_COUNT       ?= 6


This is so benchstat gives confidence intervals. Without this, it'll show the message ¹ need >= 6 samples for confidence interval at level 0.95

agilgur5 · 2024-10-26T02:17:03Z

persist/sqldb/migrate.go

+			ansiSQLChange(`drop index argo_archived_workflows_i4 on argo_archived_workflows`),
+			ansiSQLChange(`drop index argo_archived_workflows_i4`),
+		),
+		ansiSQLChange(`create index argo_archived_workflows_i4 on argo_archived_workflows (clustername, startedat)`),


how quick does this run? quick enough to include in a patch version? i.e. to cherry-pick into 3.5

Very quickly. Tested after populating my DB with 100,000 rows, all associated with the same clustername:

MySQL:

$ time go run ./hack/db migrate -d 'mysql:password@tcp/argo' INFO[0000] Migrating database schema clusterName=default dbType=mysql INFO[0000] applying database change change="drop index argo_archived_workflows_i4 on argo_archived_workflows" changeSchemaVersion=61 INFO[0000] applying database change change="create index argo_archived_workflows_i4 on argo_archived_workflows (clustername, startedat)" changeSchemaVersion=62 2024/10/26 02:24:13 Session ID: 00001 Query: create index argo_archived_workflows_i4 on argo_archived_workflows (clustername, startedat) Stack: fmt.(*pp).handleMethods@/usr/local/go/src/fmt/print.go:673 fmt.(*pp).printArg@/usr/local/go/src/fmt/print.go:756 fmt.(*pp).doPrint@/usr/local/go/src/fmt/print.go:1208 fmt.Append@/usr/local/go/src/fmt/print.go:289 log.(*Logger).Print.func1@/usr/local/go/src/log/log.go:261 log.(*Logger).output@/usr/local/go/src/log/log.go:238 log.(*Logger).Print@/usr/local/go/src/log/log.go:260 github.com/argoproj/argo-workflows/v3/persist/sqldb.ansiSQLChange.apply@/home/vscode/go/src/github.com/argoproj/argo-workflows/persist/sqldb/ansi_sql_change.go:11 github.com/argoproj/argo-workflows/v3/persist/sqldb.migrate.applyChange.func1@/home/vscode/go/src/github.com/argoproj/argo-workflows/persist/sqldb/migrate.go:301 github.com/argoproj/argo-workflows/v3/persist/sqldb.migrate.applyChange@/home/vscode/go/src/github.com/argoproj/argo-workflows/persist/sqldb/migrate.go:290 github.com/argoproj/argo-workflows/v3/persist/sqldb.migrate.Exec@/home/vscode/go/src/github.com/argoproj/argo-workflows/persist/sqldb/migrate.go:279 main.NewMigrateCommand.func1@/home/vscode/go/src/github.com/argoproj/argo-workflows/hack/db/main.go:50 github.com/spf13/cobra.(*Command).execute@/home/vscode/go/pkg/mod/github.com/spf13/[email protected]/command.go:985 github.com/spf13/cobra.(*Command).ExecuteC@/home/vscode/go/pkg/mod/github.com/spf13/[email protected]/command.go:1117 github.com/spf13/cobra.(*Command).Execute@/home/vscode/go/pkg/mod/github.com/spf13/[email protected]/command.go:1041 main.main@/home/vscode/go/src/github.com/argoproj/argo-workflows/hack/db/main.go:39 runtime.main@/usr/local/go/src/runtime/proc.go:272 runtime.goexit@/usr/local/go/src/runtime/asm_amd64.s:1700 Rows affected: 0 Last insert ID: 0 Error: upper: slow query Time taken: 0.95033s Context: context.Background real 0m1.758s user 0m1.074s sys 0m0.294s

PostgreSQL:

$ time go run ./hack/db migrate INFO[0000] Migrating database schema clusterName=default dbType=postgres INFO[0000] applying database change change="drop index argo_archived_workflows_i4" changeSchemaVersion=61 INFO[0000] applying database change change="create index argo_archived_workflows_i4 on argo_archived_workflows (clustername, startedat)" changeSchemaVersion=62 real 0m0.868s user 0m1.117s sys 0m0.271s

Dope, great work again!

agilgur5

Huge thanks for deep-diving & analyzing this, describing it, and testing it!

MasonM added 4 commits October 25, 2024 13:00

test: allow configuring benchmark iterations

fc844b3

Signed-off-by: Mason Malone <[email protected]>

fix: further optimize archived list query

3ea4b94

Signed-off-by: Mason Malone <[email protected]>

fix: optimize MySQL variant

6a90a9b

Signed-off-by: Mason Malone <[email protected]>

fix: add migration to update argo_archived_workflows_i4 index

01bbc41

Signed-off-by: Mason Malone <[email protected]>

MasonM commented Oct 26, 2024

View reviewed changes

MasonM marked this pull request as ready for review October 26, 2024 02:06

MasonM mentioned this pull request Oct 26, 2024

3.5: Further optimize Archive List API call / DB query #13601

Open

4 tasks

agilgur5 reviewed Oct 26, 2024

View reviewed changes

agilgur5 added area/workflow-archive area/api Argo Server API labels Oct 26, 2024

agilgur5 reviewed Oct 26, 2024

View reviewed changes

agilgur5 approved these changes Oct 26, 2024

View reviewed changes

agilgur5 requested a review from Joibel October 26, 2024 02:31

agilgur5 added this to the v3.5.x patches milestone Oct 26, 2024

agilgur5 requested a review from jiachengxu October 26, 2024 04:05

agilgur5 mentioned this pull request Oct 26, 2024

fix: remove JSON cast when querying archived workflows #13777

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: further optimize archive workflow listing. Fixes #13601 #13819

fix: further optimize archive workflow listing. Fixes #13601 #13819

MasonM commented Oct 26, 2024 •

edited

Loading

MasonM Oct 26, 2024

agilgur5 Oct 26, 2024 •

edited

Loading

MasonM Oct 26, 2024

agilgur5 Oct 26, 2024

agilgur5 left a comment

fix: further optimize archive workflow listing. Fixes #13601 #13819

Are you sure you want to change the base?

fix: further optimize archive workflow listing. Fixes #13601 #13819

Conversation

MasonM commented Oct 26, 2024 • edited Loading

Motivation

Modifications

Verification

MasonM Oct 26, 2024

Choose a reason for hiding this comment

agilgur5 Oct 26, 2024 • edited Loading

Choose a reason for hiding this comment

MasonM Oct 26, 2024

Choose a reason for hiding this comment

agilgur5 Oct 26, 2024

Choose a reason for hiding this comment

agilgur5 left a comment

Choose a reason for hiding this comment

MasonM commented Oct 26, 2024 •

edited

Loading

agilgur5 Oct 26, 2024 •

edited

Loading