Layers performance vs Multicompiler

[Themezv]

Hello! Could you explain the advantages and disadvantages of the approaches (multicompiler and layers features) in terms of performance?

For example. Now we are using multicompiler for 8 compilations. Almost all source files are the same – the main difference is the set of swc plugins for built-it swc loader and build targets (node, browser). Also we have slow file-system.
Can switching to layers help with speed and how?

[MILLERMARRU]

The core difference is whether you get one shared module graph or N independent ones. Multicompiler runs each config as a fully separate Compiler instance, its own resolution, its own parsing, its own module graph, even for source files that are byte-identical across all 8 of your configs. On a slow filesystem that's 8x the stat/read/resolve work for anything shared.

layers runs inside a single compilation. You tag modules with layer (via module.rules[].layer or optimization.splitChunks.cacheGroups[].layer) and then use issuerLayer to apply different rules to files depending on which layer imported them, including a different builtin:swc-loader config per layer (that's the documented pattern, issuerLayer: 'client' vs issuerLayer: 'server' with different jsc.target per rule). Since it's one compiler, one module graph, files that are shared across layers only get resolved and parsed once.

The catch for your setup specifically: layer/issuerLayer can vary loader options (so your differing swc plugin sets are a good fit), but the top-level target (node vs browser, which drives externals presets, chunk loading mechanism, global object) is still compiler-wide, it isn't something you can vary per layer. So if your 8 compilations split into, say, a node-target group and a browser-target group, you could likely collapse each group into a single layered compiler (cutting duplicate resolution/parsing within that group), but you'd still need at least 2 compiler instances total, one per actual target. Worth checking how many of your 8 configs actually share a target before deciding whether layers get you a partial or full win here.