Slide 1

Slide 1 text

Performant Android CI Mar 20th, 2025 Jason Pearson

Slide 2

Slide 2 text

Performant Android CI Mar 20th, 2025 Jason Pearson

Slide 3

Slide 3 text

Performant Android CI Jason Pearson

Slide 4

Slide 4 text

Performant Android CI Jason Pearson

Slide 5

Slide 5 text

Performant Android CI Jason Pearson The Purpose of CI πŸ’» The Journey to Performant Android CI

Slide 6

Slide 6 text

Performant Android CI Jason Pearson The Purpose of CI πŸ’» The Journey to Performant Android CI

Slide 7

Slide 7 text

Performant Android CI Build & Cache Techniques πŸ“¦ CI Pipeline Design Layered Docker Images Git Custom Checkout & Fetch

Slide 8

Slide 8 text

Performant Android CI Build & Cache Techniques πŸ“¦ CI Pipeline Design Layered Docker Images Git Custom Checkout & Fetch

Slide 9

Slide 9 text

The Purpose of CI

Slide 10

Slide 10 text

The Purpose of CI πŸ’»

Slide 11

Slide 11 text

The Purpose of CI πŸ’»

Slide 12

Slide 12 text

The Purpose of CI πŸ’»

Slide 13

Slide 13 text

πŸ’»

Slide 14

Slide 14 text

πŸ’»

Slide 15

Slide 15 text

πŸ’»

Slide 16

Slide 16 text

πŸ’»

Slide 17

Slide 17 text

πŸ’»

Slide 18

Slide 18 text

πŸ’»

Slide 19

Slide 19 text

πŸ’»

Slide 20

Slide 20 text

πŸ’»

Slide 21

Slide 21 text

πŸ’»

Slide 22

Slide 22 text

πŸ’»

Slide 23

Slide 23 text

πŸ’»

Slide 24

Slide 24 text

πŸ’» πŸ§ͺ

Slide 25

Slide 25 text

πŸ’» πŸ§ͺ

Slide 26

Slide 26 text

πŸ’» πŸ§ͺ

Slide 27

Slide 27 text

πŸ’» πŸ§ͺ

Slide 28

Slide 28 text

πŸ’» πŸ§ͺ πŸ‰

Slide 29

Slide 29 text

πŸ’» πŸ§ͺ πŸ‰

Slide 30

Slide 30 text

πŸ’» πŸ§ͺ πŸ‰

Slide 31

Slide 31 text

πŸ’» πŸ§ͺ πŸ‰

Slide 32

Slide 32 text

πŸ’» πŸ§ͺ πŸ‰

Slide 33

Slide 33 text

πŸ’» πŸ§ͺ πŸ‰

Slide 34

Slide 34 text

πŸ’» πŸ§ͺ πŸ‰

Slide 35

Slide 35 text

~900 Modules πŸ’» ~1M lines of code ~100 contributors Codebase Context Dagger Anvil (KAPT) πŸ—‘ Gradle Kotlin Scripts Kotlin 1.9

Slide 36

Slide 36 text

CI Context Single Docker image Datadog, Umbra Firebase Test Lab Gradle Develocity Self hosted EC2 Runners Artifactory GitLab CI

Slide 37

Slide 37 text

Very long MR pipeline times

Slide 38

Slide 38 text

β€’ 3-4 minutes pipeline veri fi cation jobs β€’ 30 minutes in Lint β€’ 50+ minutes in Test β€’ ~30 minutes Building APKs β€’ ~15 minutes running integration tests that are unreliable Very long MR pipeline times

Slide 39

Slide 39 text

The Journey to Performant Android CI

Slide 40

Slide 40 text

The Journey to Performant Android CI Some caching via Develocity πŸ“¦ All artifacts all the time Full git clone via GitLab Slow and flaky UI tests

Slide 41

Slide 41 text

The Journey to Performant Android CI Some caching via Develocity πŸ“¦ All artifacts all the time Full git clone via GitLab Slow and flaky UI tests

Slide 42

Slide 42 text

The Journey to Performant Android CI Everything cached πŸ“¦ All artifacts all the time Full git clone via GitLab Slow and flaky UI tests

Slide 43

Slide 43 text

The Journey to Performant Android CI Everything cached πŸ“¦ Minimize git overhead All artifacts all the time Slow and flaky UI tests

Slide 44

Slide 44 text

The Journey to Performant Android CI Everything cached πŸ“¦ Minimize git overhead Fast & Reliable UI Tests All artifacts all the time

Slide 45

Slide 45 text

The Journey to Performant Android CI Everything cached πŸ“¦ Only artifacts as needed Fast & Reliable UI Tests Minimize git overhead

Slide 46

Slide 46 text

Fast & Reliable Later… Minimize git overhead The Journey to Performant Android CI Everything cached πŸ“¦ Only artifacts as needed

Slide 47

Slide 47 text

Fast & Reliable Only artifacts as needed Minimize git overhead The Journey to Performant Android CI Everything cached πŸ“¦ Later…

Slide 48

Slide 48 text

Fast & Reliable Only artifacts as needed Minimize git overhead The Journey to Performant Android CI Everything cached πŸ“¦ Later…

Slide 49

Slide 49 text

Fast & Reliable Only artifacts as needed Minimize git overhead The Journey to Performant Android CI Later… Build & Cache Techniques πŸ“¦

Slide 50

Slide 50 text

Fast & Reliable Only artifacts as needed The Journey to Performant Android CI Later… Build & Cache Techniques πŸ“¦ Git Custom Checkout & Fetch

Slide 51

Slide 51 text

Fast & Reliable Later… CI Pipeline Design The Journey to Performant Android CI Git Custom Checkout & Fetch Build & Cache Techniques πŸ“¦

Slide 52

Slide 52 text

πŸ“¦ The Journey to Performant Android CI Build & Cache Techniques

Slide 53

Slide 53 text

Build & Cache Techniques πŸ“¦

Slide 54

Slide 54 text

JVM Memory Sizing Heap Non-Heap

Slide 55

Slide 55 text

JVM Memory Sizing Heap Non-Heap Metaspace CodeCache Young & Old Gen CompressedClassPointers GC Buffers

Slide 56

Slide 56 text

JVM Memory Sizing Heap Non-Heap Metaspace CodeCache Young & Old Gen CompressedClassPointers GC Buffers

Slide 57

Slide 57 text

JVM Memory Sizing Gradle Daemon

Slide 58

Slide 58 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon

Slide 59

Slide 59 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers

Slide 60

Slide 60 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers

Slide 61

Slide 61 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 62

Slide 62 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 63

Slide 63 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 64

Slide 64 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 65

Slide 65 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 66

Slide 66 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 67

Slide 67 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 68

Slide 68 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 69

Slide 69 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 70

Slide 70 text

JVM Memory Sizing Non-Heap memory has internal competition

Slide 71

Slide 71 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers

Slide 72

Slide 72 text

JVM Memory Sizing Gradle Daemon Kotlin Daemon Gradle Workers Remaining Memory Divided for Workers Compilation Test

Slide 73

Slide 73 text

JVM Memory Sizing Gradle Workers Compilation Test

Slide 74

Slide 74 text

JVM Memory Sizing Gradle Workers Compilation Test org.gradle.workers.max maxParallelForks 256m 512m Default Heaps:

Slide 75

Slide 75 text

The number of milliseconds to hold onto soft references per MB of heap size, default 1000 = 2h 16m 32s 8GB Heap @ 1 MS Policy 8GB Heap @ 1000 MS Policy = 0h 0m 8s 8GB Heap @ 10 MS Policy = 0h 1m 22s SoftRefLRUPolicyMSPerMB JVM Memory Sizing

Slide 76

Slide 76 text

JVM Memory Sizing

Slide 77

Slide 77 text

JVM Memory Sizing Coming Soon: Tool for Sizing

Slide 78

Slide 78 text

tcmalloc malloc replacements jemalloc Freeing up as much memory as possible. Optimize for speed while keeping memory usage steady. Pending evaluation…

Slide 79

Slide 79 text

Android Lint Best Practices πŸ‰

Slide 80

Slide 80 text

Android Lint Best Practices πŸ‰ https://googlesamples.github.io/android-custom-lint-rules/ βœ… ignoreTestSources ❌ checkAllWarnings βœ… checkDependencies Do the following:

Slide 81

Slide 81 text

Build Cache Configuration Cache Dependency Cache Script Cache Gradle Caches

Slide 82

Slide 82 text

Build Cache EXECUTING Configuration Cache Dependency Cache Script Cache Gradle Caches

Slide 83

Slide 83 text

Build Cache EXECUTING Configuration Cache Dependency Cache Script Cache Gradle Caches

Slide 84

Slide 84 text

Build Cache EXECUTING Configuration Cache Dependency Cache Script Cache Gradle Caches

Slide 85

Slide 85 text

Build Cache CONFIGURING EXECUTING Configuration Cache Dependency Cache Script Cache Gradle Caches

Slide 86

Slide 86 text

Build Cache Configuration Cache Dependency Cache Script Cache Gradle Caches

Slide 87

Slide 87 text

Gradle Build Cache

Slide 88

Slide 88 text

β€’ Reduce the number of inputs per task and make them more stable β€’ Gradle build scans to observe performance Gradle Build Cache

Slide 89

Slide 89 text

Gradle Build Cache

Slide 90

Slide 90 text

Gradle Configuration Cache on CI

Slide 91

Slide 91 text

2-3 min possible savings per job Gradle Configuration Cache on CI

Slide 92

Slide 92 text

Relies on caching outputs of configuration phase, which rely on all inputs used in the build. So we need to: 1. Find all inputs 2. Hash them 3. Search cache provider Gradle Configuration Cache on CI

Slide 93

Slide 93 text

75 60 45 30 15 0 seconds Find Grep Find w/Prune ripgrep Source Hashing Search speed πŸ”Ž

Slide 94

Slide 94 text

~/.gradle/caches/<< gradle_version >>/dependencies-accessors ~/.gradle/caches/<< gradle_version >>/generated-gradle-jars ~/.gradle/caches/<< gradle_version >>/kotlin-dsl ~/.gradle/caches/<< gradle_version >>/transforms ~/.gradle/caches/modules-2 ~/.gradle/caches/jars-9 .gradle/<< gradle_version >> .gradle/configuration-cache build-logic/**/build ci-cache.yml

Slide 95

Slide 95 text

$ aws s3 cp β€œ$(pwd)” ”$S3_BUCKET_AND_PATH" \ --exclude β€œ*" \ --include β€œ*.zip" \ --recursive AWS S3 + ZIP

Slide 96

Slide 96 text

Gradle Configuration Cache on CI Reusing configuration cache.

Slide 97

Slide 97 text

Gradle Configuration Cache on CI πŸš€

Slide 98

Slide 98 text

➜ .gradle git:(main) βœ— du -sh 8.11.1 7.4M 8.11.1 ➜ .gradle git:(main) βœ— du -sh configuration-cache 1.7M configuration-cache Gradle Configuration Cache on CI

Slide 99

Slide 99 text

➜ 8.11.1 du -sh transforms 1.1G transforms ➜ 8.11.1 du -sh generated-gradle-jars 177M generated-gradle-jars ➜ 8.11.1 du -sh kotlin-dsl 37M kotlin-dsl ➜ 8.11.1 du -sh dependencies-accessors 12M dependencies-accessors Gradle Configuration Cache on CI

Slide 100

Slide 100 text

10 8 6 4 2 0 seconds tar gzip pigz zstd lz4 Compression Options - against 1 GB πŸ—œ

Slide 101

Slide 101 text

10 8 6 4 2 0 seconds tar gzip pigz zstd lz4 Compression Options - against 1 GB πŸ—œ with multiple cores

Slide 102

Slide 102 text

10 8 6 4 2 0 seconds tar gzip pigz zstd lz4 Compression Options - against 1 GB πŸ—œ with multiple cores

Slide 103

Slide 103 text

10 8 6 4 2 0 seconds tar gzip pigz zstd lz4 Compression Options - against 1 GB πŸ—œ with multiple cores

Slide 104

Slide 104 text

10 8 6 4 2 0 seconds tar gzip pigz zstd lz4 Compression Options - against 1 GB πŸ—œ with multiple cores

Slide 105

Slide 105 text

zstd at its fastest speed has better compression than gzipped tar lz4 is lossless, which means its compression is almost nothing Multi-threaded didn’t make a difference in production Compression Options πŸ—œ

Slide 106

Slide 106 text

AWS S3 + ZSTD # find sources via ripgrep # dry run # create zstd archives # full run $ upload zstd archives via AWS S3

Slide 107

Slide 107 text

Gradle Configuration Cache on CI

Slide 108

Slide 108 text

Gradle Configuration Cache on CI Oops

Slide 109

Slide 109 text

Gradle Caches on CI πŸ€” Config Cache Dependencies Script Cache Build Cache

Slide 110

Slide 110 text

Gradle Caches on CI Config Cache Dependencies Script Cache Build Cache πŸ’‘ Task History

Slide 111

Slide 111 text

Gradle Caches on CI Config Cache Dependencies Script Cache Build Cache πŸ’‘ Task History

Slide 112

Slide 112 text

Gradle User Home in Docker COPY . /prebuild

Slide 113

Slide 113 text

Gradle User Home in Docker Custom git checkout Supporting Pipeline Layered images Track size & drift

Slide 114

Slide 114 text

Custom git checkout & fetch

Slide 115

Slide 115 text

GitLab limited options and inefficient Custom git checkout & fetch Minimize layer size

Slide 116

Slide 116 text

git clone commit tree blob

Slide 117

Slide 117 text

git clone HEAD commit tree blob

Slide 118

Slide 118 text

git clone: blobless HEAD commit tree blob

Slide 119

Slide 119 text

git clone: treeless HEAD commit tree blob

Slide 120

Slide 120 text

git clone: shallow HEAD commit tree blob

Slide 121

Slide 121 text

75 60 45 30 15 0 seconds Custom git checkout & fetch default clone blobless clone treeless clone shallow clone

Slide 122

Slide 122 text

75 60 45 30 15 0 seconds default clone blobless clone treeless clone shallow clone Custom git checkout & fetch treeless fetch

Slide 123

Slide 123 text

Custom git checkout & fetch

Slide 124

Slide 124 text

COPY . /prebuild Layered images

Slide 125

Slide 125 text

FROM zillow-docker.artifactory… COPY . /prebuild Layered images

Slide 126

Slide 126 text

FROM zillow-docker.artifactory… COPY . /prebuild RUN \ rm -rf /root/.gradle && \ mkdir -p /root/.gradle && \ mv /prebuild/gradle-user-home-backup / root/.gradle Layered images

Slide 127

Slide 127 text

FROM zillow-docker.artifactory… COPY . /prebuild RUN \ rm -rf /root/.gradle && \ mkdir -p /root/.gradle && \ mv /prebuild/gradle-user-home-backup / root/.gradle Layered images

Slide 128

Slide 128 text

# .dockerignore **/build/ Layered images

Slide 129

Slide 129 text

Layered images

Slide 130

Slide 130 text

Layered images

Slide 131

Slide 131 text

Layered images

Slide 132

Slide 132 text

Layered images Building Real Estate Debug APK Previously 30 minutes Down to 2 minutes on complete cache hit

Slide 133

Slide 133 text

πŸ§ͺ Layered images Android Unit Tests Previously 50 minutes Down to 4 minutes on complete cache hit

Slide 134

Slide 134 text

CI Pipeline Design

Slide 135

Slide 135 text

Automate building, publishing, using Docker images Flatten to reduce negative space and parallelize work Use downstream pipelines to organize and declutter CI Pipeline Design

Slide 136

Slide 136 text

CI Pipeline Design: Iterations

Slide 137

Slide 137 text

CI Pipeline Design: Iterations

Slide 138

Slide 138 text

CI Pipeline Design: Iterations

Slide 139

Slide 139 text

CI Pipeline Design: Critical Checks Group checks that always complete in 60 seconds or less. Fail entire pipeline if critical check fails, but otherwise let pipelines continue.

Slide 140

Slide 140 text

CI Pipeline Design: Critical Checks

Slide 141

Slide 141 text

CI Pipeline Design: Critical Checks

Slide 142

Slide 142 text

CI Pipeline Design: Team Jobs Dynamically run team specific work. Make it easy to keep author-team relationship up to date.

Slide 143

Slide 143 text

CI Pipeline Design: Team Jobs

Slide 144

Slide 144 text

CI Pipeline Design: Today

Slide 145

Slide 145 text

CI Pipeline Design: Today

Slide 146

Slide 146 text

CI Pipeline Design: Today

Slide 147

Slide 147 text

CI Pipeline Design: Today

Slide 148

Slide 148 text

CI Pipeline Design: Today

Slide 149

Slide 149 text

CI Pipeline Design: Downstream Included Build Gradle Code Analysis Unit Tests

Slide 150

Slide 150 text

Observability - Docker Overhead

Slide 151

Slide 151 text

Observability - Docker Overhead

Slide 152

Slide 152 text

3 2 1 0 minutes Original Gradle Tools Observability - Docker Overhead

Slide 153

Slide 153 text

3 2 1 0 minutes Original Gradle Tools Increased pull time nullified with pre-pulling images Observability - Docker Overhead

Slide 154

Slide 154 text

3 2 1 0 minutes Original Gradle Tools Increased pull time nullified with pre-pulling images Observability - Docker Overhead

Slide 155

Slide 155 text

Observability - Code Drift git rev-list --left-right \ --count β€œ${OLD_SHA}...${NEW_SHA}"

Slide 156

Slide 156 text

Observability - Code Drift

Slide 157

Slide 157 text

No content

Slide 158

Slide 158 text

So how fast is it today?

Slide 159

Slide 159 text

p75 30 min

Slide 160

Slide 160 text

Observability - Code Drift

Slide 161

Slide 161 text

The Journey to Performant Android CI

Slide 162

Slide 162 text

What’s next? Emulator.wtf Local Build Perf Upgrading to Kotlin 2 πŸ’»

Slide 163

Slide 163 text

Jason Pearson Questions?