Currently this repository hosts two kinds of benchmarks:
1. The older "milli benchmarks", that use [criterion](https://github.com/bheisler/criterion.rs) and live in the "benchmarks" directory.
2. The newer "bench" that are workload-based and so split between the [`workloads`](./workloads/) directory and the [`xtask::bench`](./xtask/src/bench/) module.
This document describes the newer "bench" benchmarks. For more details on the "milli benchmarks", see [benchmarks/README.md](./benchmarks/README.md).
## Design philosophy for the benchmarks
The newer "bench" benchmarks are **integration** benchmarks, in the sense that they spawn an actual Meilisearch server and measure its performance end-to-end, including HTTP request overhead.
Since this is prone to fluctuating, the benchmarks regain a bit of precision by measuring the runtime of the individual spans using the [logging machinery](./CONTRIBUTING.md#logging) of Meilisearch.
A span roughly translates to a function call. The benchmark runner collects all the spans by name using the [logs route](https://github.com/orgs/meilisearch/discussions/721) and sums their runtime. The processed results are then sent to the [benchmark dashboard](https://bench.meilisearch.dev), which is in charge of storing and presenting the data.
## Running the benchmarks
Benchmarks can run locally or in CI.
### Locally
#### With a local benchmark dashboard
The benchmarks dashboard lives in its [own repository](https://github.com/meilisearch/benchboard). We provide binaries for Ubuntu/Debian, but you can build from source for other platforms (MacOS should work as it was developed under that platform).
Run the `benchboard` binary to create a fresh database of results. By default it will serve the results and the API to gather results on `http://localhost:9001`.
From the Meilisearch repository, you can then run benchmarks with:
For processing the results, look at [Looking at benchmark results/Without dashboard](#without-dashboard).
### In CI
We have dedicated runners to run workloads on CI. Currently, there are three ways of running the CI:
1. Automatically, on every push to `main`.
2. Manually, by clicking the [`Run workflow`](https://github.com/meilisearch/meilisearch/actions/workflows/bench-manual.yml) button and specifying the target reference (tag, commit or branch) as well as one or multiple workloads to run. The workloads must exist in the Meilisearch repository (conventionally, in the [`workloads`](./workloads/) directory) on the target reference. Globbing (e.g., `workloads/*.json`) works.
3. Manually on a PR, by posting a comment containing a `/bench` command, followed by one or multiple workloads to run. Globbing works. The workloads must exist in the Meilisearch repository in the branch of the PR.
```
/bench workloads/movies*.json /hackernews_1M.json
```
## Looking at benchmark results
### On the dashboard
Results are available on the global dashboard used by CI at <https://bench.meilisearch.dev> or on your [local dashboard](#with-a-local-benchmark-dashboard).
The dashboard homepage presents three sections:
1. The latest invocations (a call to `cargo xtask bench`, either local or by CI) with their reason (generally set to some helpful link in CI) and their status.
2. The latest workloads ran on `main`.
3. The latest workloads ran on other references.
By default, the workload shows the total runtime delta with the latest applicable commit on `main`. The latest applicable commit is the latest commit for workload invocations that do not originate on `main`, and the latest previous commit for workload invocations that originate on `main`.
You can explicitly request a detailed comparison by span with the `main` branch, the branch or origin, or any previous commit, by clicking the links at the bottom of the workload invocation.
In the detailed comparison view, the spans are sorted by improvements, regressions, stable (no statistically significant change) and unstable (the span runtime is comparable to its standard deviation).
You can click on the name of any span to get a box plot comparing the target commit with multiple commits of the selected branch.
### Without dashboard
After the workloads are done running, the reports will live in the Meilisearch repository, in the `bench/reports` directory (by default).
You can then convert these reports into other formats.
- To [Firefox profiler](https://profiler.firefox.com) format. Run:
```sh
cd bench/reports
cargo run --release --bin trace-to-firefox -- my_workload_1-0-trace.json
```
You can then upload the resulting `firefox-my_workload_1-0-trace.json` file to the online profiler.
## Designing benchmark workloads
Benchmark workloads conventionally live in the `workloads` directory of the Meilisearch repository.
They are JSON files with the following structure (comments are not actually supported, to make your own, remove them or copy some existing workload file):
```jsonc
{
// Name of the workload. Must be unique to the workload, as it will be used to group results on the dashboard.
"name": "hackernews.ndjson_1M,no-threads",
// Number of consecutive runs of the commands that should be performed.
// Each run uses a fresh instance of Meilisearch and a fresh database.
// Each run produces its own report file.
"run_count": 3,
// List of arguments to add to the Meilisearch command line.
// An expression that can be parsed as a comma-separated list of targets and levels
// as described in [tracing_subscriber's documentation](https://docs.rs/tracing-subscriber/latest/tracing_subscriber/filter/targets/struct.Targets.html#examples).
// The expression is used to filter the spans that are measured for profiling purposes.
// Optional, defaults to "indexing::=trace" (for indexing workloads), common other values is
// When using "asset", use the name of an asset as value to use the content of that asset as body.
// the content type is derived of the format of the asset:
// "NdJson" => "application/x-ndjson"
// "Json" => "application/json"
// "Raw" => "application/octet-stream"
// See [AssetFormat::to_content_type](https://github.com/meilisearch/meilisearch/blob/7b670a4afadb132ac4a01b6403108700501a391d/xtask/src/bench/assets.rs#L30)
