Dockerfile.driver

# syntax=docker/dockerfile:1.7-labs
#
# The fleet simulator (`fleet-sim`) as the driver of an *attached* fleet: it owns no replicas and no
# filesystem, it attaches to ones somebody else is running. Built from the workspace root:
#
#     docker build -f Dockerfile.driver -t fleetsim-driver .
#
# Companion to `Dockerfile.agent`, which builds the replicas this drives.
#
# ## Why this image exists at all
#
# The checker reads the shards through the *local* filesystem (`check.rs`), so the driver has to be
# somewhere the shards are mounted — inside the VPC, with the same EFS. It also hosts the mock model
# server and the exec double the replicas dial back into, at a fixed address the replicas were
# started with. Both of those make it a task rather than something run from a laptop.
#
# The AWS CLI is here for one reason: `--fault-cmd`. The simulator does not own these replicas, so
# `kill`, `term`, `partition` and `heal` are somebody else's to perform — on AWS, the EC2 and ECS
# APIs. The CLI stays in *this* image and out of the agent image, and the words `ECS`/`EFS`/`AWS`
# stay out of the Rust entirely.

ARG RUST_VERSION=1
ARG ALPINE_VERSION=3.22

FROM rust:${RUST_VERSION}-bookworm AS chef
ARG TARGETARCH
RUN apt-get update && apt-get install -y --no-install-recommends musl-tools \
    && rm -rf /var/lib/apt/lists/*
RUN case "$TARGETARCH" in \
      amd64) echo x86_64-unknown-linux-musl ;; \
      arm64) echo aarch64-unknown-linux-musl ;; \
      *) echo "unsupported TARGETARCH=$TARGETARCH" >&2; exit 1 ;; \
    esac > /etc/cargo-target \
    && rustup target add "$(cat /etc/cargo-target)"
RUN cargo install cargo-chef --locked --version ^0.1
WORKDIR /app

# Same flag `Dockerfile.agent` sets, and here for a blunter reason than its RSS one: the default
# musl target links `crt-static` as a **static-PIE**, and the `fleet-sim` binary built that way
# segfaults before `main` in this base image — silently, exit 139, with no output at all, which from
# the outside looks exactly like a matrix that started and hung. The agent image has always carried
# this flag and has always run here, so the driver carries it too rather than being the one binary
# built a way nothing else in this repo ships.
ENV RUSTFLAGS="-C relocation-model=static"

FROM chef AS planner
COPY . .
RUN cargo chef prepare --recipe-path recipe.json

FROM chef AS builder
COPY --from=planner /app/recipe.json recipe.json
RUN --mount=type=cache,target=/usr/local/cargo/registry \
    --mount=type=cache,target=/usr/local/cargo/git \
    target="$(cat /etc/cargo-target)" \
    && linker_env="CARGO_TARGET_$(echo "$target" | tr 'a-z-' 'A-Z_')_LINKER" \
    && cc_env="CC_$(echo "$target" | tr '-' '_')" \
    && export "$linker_env=musl-gcc" "$cc_env=musl-gcc" \
    && cargo chef cook --release --target "$target" -p beyond-ai-fleet-sim --recipe-path recipe.json
COPY . .
RUN --mount=type=cache,target=/usr/local/cargo/registry \
    --mount=type=cache,target=/usr/local/cargo/git \
    target="$(cat /etc/cargo-target)" \
    && linker_env="CARGO_TARGET_$(echo "$target" | tr 'a-z-' 'A-Z_')_LINKER" \
    && cc_env="CC_$(echo "$target" | tr '-' '_')" \
    && export "$linker_env=musl-gcc" "$cc_env=musl-gcc" \
    && cargo build --release --target "$target" -p beyond-ai-fleet-sim --bin fleet-sim \
    && cp "target/$target/release/fleet-sim" /usr/local/bin/fleet-sim

# The metrics sidecar, in its own image. Nothing but socat, because it is pulled by every replica
# task and the driver image below is ~400 MB of AWS CLI that a port forwarder has no use for.
#
#     docker build -f Dockerfile.driver --target socat -t fleetsim-socat .
FROM alpine:${ALPINE_VERSION} AS socat
RUN apk add --no-cache socat
ENTRYPOINT ["socat"]

# ---------------------------------------------------------------------------
# The driver: Debian, not Alpine, and the reason is one binary.
#
# `aws ecs execute-command` needs the Session Manager plugin, which is how the `kill` fault gets a
# SIGKILL into a replica — the only way to reach inside a Fargate task, since ECS has no
# "signal this container" API and StopTask always sends SIGTERM first. AWS ships that plugin as a
# `.deb` and an `.rpm` against glibc and publishes nothing for musl, so on Alpine the fault failed
# with `SessionManagerPlugin is not found` — swallowed, because the call is best-effort, and
# surfacing 250 s later as "r1 did not reach STOPPED".
#
# The alternative was to make `kill` mean something else (partition from storage, then StopTask),
# which would have produced a replica that died without sealing — the right *outcome* — while no
# longer being the same fault the local substrate performs with `kill -9`. A claim proved one way
# on the homelab and another way here is two claims, so the image changes instead of the fault.
#
# `fleet-sim` is a static musl binary and does not care which libc is installed.
# ---------------------------------------------------------------------------
FROM debian:stable-slim AS runtime
RUN apt-get update && apt-get install -y --no-install-recommends \
      ca-certificates curl unzip jq socat less procps \
    && rm -rf /var/lib/apt/lists/*
ARG TARGETARCH
RUN set -eux; \
    case "$TARGETARCH" in \
      amd64) awscli=x86_64; plugin=64bit ;; \
      arm64) awscli=aarch64; plugin=arm64 ;; \
      *) echo "unsupported TARGETARCH=$TARGETARCH" >&2; exit 1 ;; \
    esac; \
    curl -sSL "https://awscli.amazonaws.com/awscli-exe-linux-${awscli}.zip" -o /tmp/awscli.zip; \
    unzip -q /tmp/awscli.zip -d /tmp; /tmp/aws/install; \
    curl -sSL "https://s3.amazonaws.com/session-manager-downloads/plugin/latest/ubuntu_${plugin}/session-manager-plugin.deb" \
      -o /tmp/smp.deb; \
    dpkg -i /tmp/smp.deb; \
    rm -rf /tmp/awscli.zip /tmp/aws /tmp/smp.deb; \
    aws --version; session-manager-plugin --version

COPY --from=builder /usr/local/bin/fleet-sim /usr/local/bin/fleet-sim
COPY deploy/efs-proof/fault.sh /usr/local/bin/fault.sh
RUN chmod 0755 /usr/local/bin/fault.sh

# Root on purpose. The EFS access point pins every operation to uid/gid 10001 regardless of the
# caller, so the container user buys no isolation here — and the driver has to read every tenant's
# shard, which is exactly what the access point already grants it.
WORKDIR /mnt/efs
ENTRYPOINT ["/bin/bash"]