biome is very slow when run inside container compared to native (~10x)

[ulrichbaur]

When trying to integrate this docker image for use in Gitlab CI, I noticed that biome takes about 10x longer to execute when using this image compared to running biome natively. This is reproducable outside of Gitlab CI and the reason seems to be that the image is based on alpine. My assumption is that it has something to do with alpine using musl instead of glibc.

When I swap out the base image to debian:slim , the performance is very close to native. For additional context, all the systems I'm working with are using x64, so I haven't had the chance to try it out on other architectures.

FROM debian:trixie-slim AS fetcher

ARG ARCH="x64"
ARG BIOME_VERSION=1.9.4

ENV BIOME_VERSION=${BIOME_VERSION}

# biome v1
ADD https://github.com/biomejs/biome/releases/download/cli/v${BIOME_VERSION}/biome-linux-${ARCH} /usr/local/bin/biome

# biome v2
# ADD https://github.com/biomejs/biome/releases/download/@biomejs/biome@${BIOME_VERSION}/biome-linux-${ARCH} /usr/local/bin/biome

RUN chmod +x /usr/local/bin/biome

FROM debian:trixie-slim AS biome

WORKDIR /code

RUN apt update && apt install -y git && rm -rf /var/lib/apt/lists/* && git config --global --add safe.directory /code
COPY --from=fetcher /usr/local/bin/biome /usr/local/bin/biome

ENTRYPOINT [ "/usr/local/bin/biome" ]

[ematipico]

Do you have a solution in mind?

[ulrichbaur]

What worked for me was switching the base image to debian-slim like in the Dockerfile I provided above. Since I don't know for what reasons alpine was chosen, I don't know if my "fix" works for everyone. Hence I made an issue here first.

As an example, with the official image it took ~6s to format our repository and with the image based on the Dockerfile above ~550 ms (basically indistinguishable from running biome without docker container).

[nhedger]

Did a quick benchmark which seems to agree with you.

hyperfine --warmup 3 \
  'docker run -v $(pwd):/code biome-alpine-musl check' \
  'docker run -v $(pwd):/code biome-debian-glibc check' \
  'docker run -v $(pwd):/code biome-debian-musl check'

Benchmark 1: docker run -v $(pwd):/code biome-alpine-musl check
  Time (mean ± σ):      1.121 s ±  0.019 s    [User: 0.019 s, System: 0.013 s]
  Range (min … max):    1.080 s …  1.146 s    10 runs

Benchmark 2: docker run -v $(pwd):/code biome-debian-glibc check
  Time (mean ± σ):     374.0 ms ±  15.0 ms    [User: 19.1 ms, System: 13.8 ms]
  Range (min … max):   348.9 ms … 394.8 ms    10 runs

Benchmark 3: docker run -v $(pwd):/code biome-debian-musl check
  Time (mean ± σ):      1.149 s ±  0.027 s    [User: 0.019 s, System: 0.014 s]
  Range (min … max):    1.119 s …  1.206 s    10 runs

Summary
  docker run -v $(pwd):/code biome-debian-glibc check ran
    3.00 ± 0.13 times faster than docker run -v $(pwd):/code biome-alpine-musl check
    3.07 ± 0.14 times faster than docker run -v $(pwd):/code biome-debian-musl check

This might not even have anything to do with Docker btw, but we should probably profile musl and glibc binaries to see what is happening

[nhedger]

For most targets, we use tikv-jemallocator as the global allocator, however, for musl builds we fall back to the default allocator for compat reasons, which is a bit slower.

I'm exploring the possibility of using mimalloc for musl builds, as we do for Windows builds because it has a performance comparable to jemallocator