COMP: Fix ccache hit rate on Azure DevOps to near-zero

[hjmjohnson]

Fix ccache configuration on all 7 Azure DevOps pipelines — hit rate was 0.02%, now 99.95%.

Root cause and fix

Root cause: CCACHE_NODIRECT=1 (global variable) disabled ccache's fast "direct mode" which hashes source files directly. Instead, every file was preprocessed before hashing — a slower path that also changes hash inputs, causing near-total cache misses across runs.

Evidence: ARM CI (GitHub Actions) never had NODIRECT and achieves 98.5% hit rate. Azure DevOps with NODIRECT=1 was at 0.02% (1 hit out of 5160 cacheable calls).

Fixes applied to all 7 Azure pipelines:

1. Remove CCACHE_NODIRECT=1 — enable direct mode
2. Add CCACHE_SLOPPINESS=pch_defines,time_macros — prevent __DATE__/__TIME__ misses
3. Add per-job name to cache key (ccache-v4 | OS | JobName | SHA) — prevents jobs with different build configs from poisoning each other's caches
4. Remove ITK_USE_CCACHE:BOOL=ON — redundant with CMAKE_*_COMPILER_LAUNCHER=ccache
5. Add ccache --show-stats + ccache --zero-stats steps for monitoring

Pipelines updated:

• AzurePipelinesLinux.yml (3 jobs: Linux, LinuxLegacyRemoved, LinuxCxx20)
• AzurePipelinesLinuxPython.yml
• AzurePipelinesMacOS.yml
• AzurePipelinesMacOSPython.yml
• AzurePipelinesWindows.yml
• AzurePipelinesWindowsPython.yml
• AzurePipelinesBatch.yml

Test results from PR #6047

Metric	Before fix	Run 1 (cold)	Run 2 (warm)
Hit rate	0.02%	0% (seeding)	99.95%
Direct hits	1	0	2166/2167
Misses	5159	~2167	1

[greptile-apps]

Greptile Summary

This PR fixes a near-zero ccache hit rate (0.02%) on the Azure DevOps Linux pipeline by removing CCACHE_NODIRECT=1, which was forcing ccache into preprocessor mode and causing near-total cache misses across builds. Additional improvements include per-job cache key isolation, CCACHE_SLOPPINESS=pch_defines,time_macros to suppress __DATE__/__TIME__ invalidation, and ccache stats steps for ongoing monitoring.

Confidence Score: 5/5

Safe to merge — CI-only configuration change with well-evidenced improvement and no P0/P1 findings.

All changes are confined to CI YAML configuration. The root cause analysis is thorough and supported by measured data (0.02% → 99.95% hit rate). Each change is individually justified: removing CCACHE_NODIRECT restores direct mode, per-job cache keys prevent cross-configuration pollution, and the stats steps add observability. The only note is the deliberate time_macros sloppiness trade-off, which is P2 and standard practice for CI caching.

No files require special attention.

Important Files Changed

Filename	Overview
Testing/ContinuousIntegration/AzurePipelinesLinux.yml	Replaces CCACHE_NODIRECT=1 with CCACHE_SLOPPINESS=pch_defines,time_macros, adds per-job cache keys (ccache-v4

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Pipeline starts per job] --> B[Restore ccache per-job key]
    B --> C[zero-stats, evict old entries, show-config]
    C --> D[ctest build with ccache direct mode]
    D --> E{Cache hit?}
    E -->|Yes ~99.95pct| F[Reuse cached object]
    E -->|No| G[Compile and store]
    D --> H[show-stats always runs]
    H --> I[report diagnostics and publish]

Loading

_{Reviews (1): Last reviewed commit: "COMP: Fix near-zero ccache hit rate on A..." | Re-trigger Greptile}