> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bluejutzu.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# GLua Performance

> Measured latency and memory on a 932-file, 232,000-line Garry's Mod gamemode, and how the server keeps those numbers where they are.

Every number below was measured against a real gamemode: 932 files, 232,000 lines. Nothing on this page is a target — it is what the current build does on that codebase.

## Benchmarks

| Metric                     | Result                                |
| -------------------------- | ------------------------------------- |
| Cold index                 | 4.8 s (about 48,000 lines per second) |
| Retained heap              | 59 MB                                 |
| Re-analyse 3,200-line file | 31 ms                                 |
| Completion (member)        | 1.2 ms                                |
| Completion (global scope)  | 4.1 ms                                |
| Diagnostics                | 3.5 ms                                |

## How the numbers stay where they are

* **Batched indexing.** The initial scan runs in chunks and yields between them, so the editor is usable while the workspace is still loading rather than blocked until it finishes.
* **Syntax trees are dropped once they have been read.** A closed file keeps only the facts extracted from it — globals it defines, hooks it adds, net messages it sends — and releases the tree along with every closure that captured it. On the same 232,000-line project the trees alone were 770 MB; keeping only the facts brings retained heap to 59.
* **Everything cross-file is stored as data, not trees.** The symbol index, the net message pairing, the call graph behind [hot paths](/glua/features/hot-paths) and call hierarchy — extracted once while a tree is in hand and left alone if the file's contents did not change. A scan of the whole project does not rebuild them one file at a time.
* **Call edges are resolved when something asks for them.** In a workspace where hundreds of files define the same path, indexing every edge up front is quadratic in the number of definitions, and almost none of those edges are ever looked at.

## Running your own benchmark

Point the bench at any addon or gamemode on disk:

```bash theme={"system"}
pnpm run bench -- path/to/addon
```

You get the cold index time and retained heap, per-feature latency for completion, diagnostics and hover, and every diagnostic grouped by rule with an example of each. The last one is the useful one: a change that makes the output noisier shows up as a rule whose count jumped, with a line to look at.

## Known limits

* **Type inference is shallow on purpose.** It answers "what can follow this dot" and stays quiet where deeper inference would be expensive. Complex generics and heavily metatable-driven code need an explicit `---@type` or `---@param`.
* **A file that changes is re-analysed whole.** At 31 ms for a 3,200-line file, behind a 250 ms debounce, incremental reparsing is not worth what it would cost in complexity.
* **The wiki is the source of truth.** If an entry is missing or a signature is wrong on the wiki, it is missing or wrong here. Rebuild the dataset after a Garry's Mod update.