Fix: Unity Asset Bundles Rebuild Every Time Incremental

CI builds 600 asset bundles every time even when nothing changed. Build wall-clock: 22 minutes. The fix is to stop deleting the output and to switch to SBP for fine control.

The Symptom

Bundle build takes the full duration each time. CI logs show every bundle being processed. Cache hit rate (visible in Console with verbose logging) is near zero.

What Causes This

Legacy BuildPipeline.BuildAssetBundles uses the output directory’s previous .manifest files to determine what already-built bundles can be reused. If you wipe the directory before building, every bundle is “new” and Unity rebuilds everything.

The Library/AssetBundleCache cache also helps but only when the build pipeline trusts the prior output.

The Fix

Step 1: Preserve output between builds.

// Bad — rebuilds everything
Directory.Delete(outputPath, recursive: true);
Directory.CreateDirectory(outputPath);
BuildPipeline.BuildAssetBundles(outputPath, BuildAssetBundleOptions.None, target);

// Good — uses cache
if (!Directory.Exists(outputPath))
    Directory.CreateDirectory(outputPath);
BuildPipeline.BuildAssetBundles(outputPath, BuildAssetBundleOptions.None, target);

Step 2: Cache the Library/ folder in CI. Library/AssetBundleCache holds intermediate hash data. Cache it across CI runs (GitHub Actions cache, Fly volumes, etc.) for 5–10x faster builds on consecutive commits.

Step 3: Switch to Scriptable Build Pipeline. Add com.unity.scriptablebuildpipeline via Package Manager. Use ContentPipeline.BuildAssetBundles instead.

var input = ContentPipeline.GenerateBundleBuilds(...);
var parameters = new BundleBuildParameters(target, group, outputPath);
var result = ContentPipeline.BuildAssetBundles(parameters, new BundleBuildContent(input.results), out var results);

SBP gives you proper incremental builds, deterministic outputs (same content => same bytes), and CacheServer integration.

Hash Stability

Bundle hashes are computed from referenced asset bytes, dependencies, and the bundle name. Renaming a referenced texture changes the bundle hash even though the texture bytes are the same. Be careful with case sensitivity (Foo.png vs foo.png) on cross-platform CI — Linux is case-sensitive and may produce different hashes than your Mac/PC dev box.

Verifying

Build, change one file, build again. Console output (verbose) should mention “Bundle X up to date.” Time should be a fraction of the full build. If still rebuilding everything, the cache or the output isn’t being preserved.

Understanding the issue

Build pipelines transform development assets into shipping packages. Each transformation can introduce subtle changes: compression, stripping, format conversion, code generation. A bug that only appears in the cooked build is usually one of these transformations doing something the author didn't expect.

The specific bug described above is the kind that surfaces during integration rather than unit testing. It depends on a combination of factors: the asset configuration, the runtime state, the platform's specific behavior. In isolation, each piece looks correct; in combination, the bug emerges. This is why thorough integration testing - playing the actual game in realistic conditions - catches things that automated tests miss.

Why this happens

This bug class disproportionately affects late-stage development. The work to surface it is interactive testing in realistic conditions, which only really happens after the gameplay is in place and assets are populated. Catching it early requires deliberate testing of conditions that look unimportant.

At the engine level, the behavior comes from a deliberate design decision in Unity. The engine team chose a particular trade-off - usually performance versus convenience, or generality versus specificity - and that trade-off has consequences when you push against it. Understanding the trade-off is what turns 'this bug is mysterious' into 'this bug is the expected consequence of this design'.

Verifying the fix

After applying the fix, the verification step has three parts: confirm the original repro is resolved, confirm no obvious regressions in adjacent functionality, and (for shipping titles) deploy to a small player cohort first and watch the crash and report rates. Each step catches something the others miss.

Reproducibility is the prerequisite for verification. If you can't reliably reproduce the bug pre-fix, you can't reliably verify it post-fix. Spend time getting a clean reproduction before you write any fix code. The fix is fast once you understand the reproduction; the reproduction is the slow part.

Variations to watch for

There's almost always a less obvious case where the same problem applies. The reported case is the one a player hit; the related cases hide because they're rarer or affect fewer players. After fixing the reported case, search the codebase for the pattern - one fix often unlocks several.

Adjacent bugs often share a root cause. After fixing the case you've found, spend an hour searching the codebase for similar patterns. What's the same call with different arguments? The same data flow with a different entity type? The same lifecycle issue in a sibling system? Each match is a candidate for the same fix, or a related fix that prevents future bugs of the same class.

In production

For shipping titles with a long support window, watch for this issue resurfacing after dependency updates. Engine upgrades, driver updates, OS releases - each one can resurface a bug class you thought you'd fixed because the underlying behavior changed slightly. Regression tests catch the obvious ones; player reports catch the rest.

When triaging a similar issue in production, prioritize gathering data over hypothesizing causes. A player report describes a symptom; what you need is a build SHA, a session timestamp, and ideally a screen recording or session replay. With those, the bug becomes tractable. Without them, you're guessing at hypothetical reproductions that may not match what the player actually hit.

Performance considerations

Performance implications matter when this bug class scales with player count or asset count. A bug that fires once per session is annoying; a bug that fires once per frame compounds. After fixing, profile the affected code path under realistic load. The fix that's correct for one entity may be too slow for ten thousand.

Diagnostic approach

Diagnosing this class of bug benefits from a structured approach: confirm the symptom, isolate the variables, hypothesize the cause, and verify the hypothesis before writing fix code. Skipping the isolation step is the most common mistake; without it, fixes often address symptoms while the underlying cause continues to produce other variations.

For Unity-specific diagnostics, the editor's profiler is the canonical starting point. Capture a representative frame with the symptom present; compare against a frame without the symptom; the diff often points directly at the cause. If the symptom is non-deterministic, capture multiple frames and look for the pattern - the cause is usually a state transition or a specific input value rather than a continuous effect.

Tooling and ecosystem

Modern engine versions ship better tooling for this kind of issue than older versions. If you're on an older release, the diagnostic step may take significantly longer because the tools you'd want don't exist yet. Sometimes the right answer is upgrading rather than fighting through limited tooling.

Within Unity, the relevant diagnostic surfaces include the standard frame debugger, memory profiler, and engine-specific debug overlays. Each one shows a different facet of what's happening. The frame debugger reveals draw call ordering and state transitions; the memory profiler shows allocation patterns; the debug overlay reveals per-system state. Bugs that resist one tool usually surrender to another - the trick is knowing which tool to reach for first.

Edge cases and pitfalls

Boundary conditions deserve specific testing attention. What happens when the input is zero, maximum, negative, or NaN? What happens at the start of a session vs hours in? What happens at the boundary between two systems handling the same data? These are where bugs hide and where regression tests are most valuable.

When writing a regression test for this fix, focus on the boundary conditions that surfaced the original bug. Tests that exercise the happy path catch obvious regressions; tests that exercise the boundary catch the subtler regressions that look like new bugs but are really the original returning. The latter are the tests that earn their keep over the long life of the project.

Team communication

When this bug class affects multiple teams (often the case for cross-system issues), early communication prevents duplicate work. The team that owns the symptom may not own the cause. A 15-minute conversation at the start of triage often saves hours of independent investigation.

If this fix touches a system several engineers work in, a short writeup in the team's engineering channel helps. Not a full design doc - a paragraph explaining what was wrong, what's fixed, and what to watch for. Future engineers encountering similar symptoms will search for the fix; making it findable is a small investment that pays back later.

“Don’t wipe output. Cache the Library. SBP for real control. Builds get fast.”

Related Issues

For Addressables handle leak, see handle leak. For Android NDK errors, see NDK errors.

Preserve output. Cache Library. SBP. Fast builds.