Read 2026-10-22
ray-project/ray: technical due diligence, read from its repository
What Borehole read in ray-project/ray's code and history at e0546a4: its licence, security advisories, releases, review practice, tests, dependencies and operations. Every finding cites the file and line it rests on. The full report also covers the team and asks the questions the code cannot answer.
| Dimension | Rating |
|---|---|
| Architecture & Codebase — including AI provenance | adequate |
| Engineering Process & SDLC Maturity | strong |
| Product & Engineering Maturity | strong |
| Security & Reliability Posture | adequate |
| Observability & Operations | adequate |
| Fit with the Acquirer | not assessable |
What the repository says
- Releases are cut from another branch than the one surveyed
AGENTS.md:3: states a policy on AI-generated contributions
Architecture & Codebase — including AI provenance
adequatepython/ray/data/dataset.py is 8,940 lines, 70× this codebase's median file
adequateThe median source file here is 127 lines. 88 files exceed 1,270 lines, the largest being 8,940. Size is not a defect on its own — some problems genuinely live in one place — but these are where merge conflicts, review fatigue and single-owner knowledge concentrate, and they are the first thing that slows an inheriting engineer down.
python/ray/data/dataset.py:1 — 8,940 lines (70× median) |
python/ray/serve/_private/deployment_state.py:1 — 7,702 lines (60× median) |
python/ray/data/read_api.py:1 — 6,780 lines (53× median) |
python/ray/rllib/algorithms/algorithm_config.py:1 — 6,631 lines (52× median) |
python/ray/_raylet.pyx:1 — 5,538 lines (43× median) |
2 base images float: no version, or `latest`
adequate2 of 9 `FROM` lines in product Dockerfiles name an image without a version, or with `latest`. Each build pulls whatever the registry serves that day, so the image that ships can change without a commit, and last month's build cannot be reproduced. Pin a version, or a digest.
docker/ray-llm/Dockerfile:57 — base image without a version |
docker/ray-llm/Dockerfile:112 — base image without a version |
Engineering Process & SDLC Maturity
strongNo finding.
Product & Engineering Maturity
strong38 files still carry code deprecated more than two years ago
adequate38 files were marked deprecated over two years ago and still are, the oldest since Apr 2022. A deprecation is a promise to remove. Ask what still depends on these, and whether removing them is scheduled.
python/ray/tune/tune.py — deprecated since Apr 2022 |
python/ray/data/read_api.py — deprecated since May 2022 |
python/ray/tune/resources.py — deprecated since May 2022 |
python/ray/train/torch/train_loop_utils.py — deprecated since Jun 2022 |
python/ray/util/client_connect.py — deprecated since Jun 2022 |
Releases are cut from another branch than the one surveyed
adequateThe newest release, ray-2.59.0, is not on the branch this survey read; the newest release on it is ray-0.7.0. The shipped product may differ from what this report describes. Survey the release branch too.
| tag ray-2.59.0 |
Security & Reliability Posture
adequate5 Dockerfiles run the application as root
adequateNo USER instruction switches away from root, so the process runs with full privileges inside the container. It is a one-line fix and an ordinary expectation, which is why its absence is worth asking about: it usually says more about whether anyone has reviewed the deployment than about this container specifically.
docker/base-extra/Dockerfile — runs as root |
docker/ray-llm/Dockerfile — runs as root |
docker/ray-ml/Dockerfile — runs as root |
docker/ray-torch/Dockerfile — runs as root |
1 environment file is committed
adequatepython/ray/dashboard/client/.env.production.local is tracked in git. Environment files are where credentials usually live, and committing one usually means nobody decided how secrets reach the running application. Whatever is in it is in the history for good. Check 05.1 says whether it holds anything credential-shaped.
python/ray/dashboard/client/.env.production.local — committed environment file |
The build runs third-party code it does not pin
adequate7 steps pipe a download straight into a shell. Tags have been moved to malicious commits in real supply-chain attacks, and a piped script is whatever the server returns that minute. Pinning to a commit is the fix, and it is small.
.buildkite/cicd.rayci.yml:8 — download piped into a shell |
28 outbound calls can wait forever
adequate28 HTTP calls use a client with no timeout (Python `requests` without `timeout=`, or Go's default client). When a dependency slows down, each one holds a worker until the connection gives up, and enough of them take the service down with the dependency.
python/ray/autoscaler/_private/spark/node_provider.py:77 — outbound call without a timeout |
python/ray/autoscaler/_private/spark/node_provider.py:134 — outbound call without a timeout |
python/ray/autoscaler/v2/instance_manager/cloud_providers/kuberay/ippr_provider.py:677 — outbound call without a timeout |
17 shell commands built from strings
adequate17 places run a command through a shell (`shell=True`, `os.system`, `exec` with a template) rather than passing arguments. Where any part comes from outside, it is command injection. Ask where each one's input comes from.
python/ray/_private/runtime_env/context.py:130 — command run through a shell |
python/ray/autoscaler/_private/fake_multi_node/command_runner.py:30 — command run through a shell |
python/ray/autoscaler/_private/fake_multi_node/node_provider.py:476 — command run through a shell |
66 deserialisers that can run code
adequate66 places read data with a deserialiser that executes what it reads (pickle, `yaml.load` without a safe loader). That is safe on data the system wrote itself and remote code execution on anything else. Ask where each one's input comes from.
python/ray/_private/function_manager.py:278 — code-executing deserialiser |
python/ray/_private/function_manager.py:314 — code-executing deserialiser |
python/ray/_private/function_manager.py:683 — code-executing deserialiser |
44 locked dependencies have published vulnerabilities, 70 high or critical in what ships
adequate44 of 908 locked dependencies match 121 published advisories in OSV, 70 of them rated high or critical. A lockfile says which version is installed, not whether the flawed code is reachable from this product, so each needs a look. Most are fixed by an upgrade, and the advisory names the version.
python/ray/dashboard/client/package-lock.json — form-data 3.0.1: GHSA-fjxv-7rqg-78g4 (critical), fixed in 3.0.4 |
python/ray/dashboard/client/package-lock.json — proxy-addr 2.0.7: GHSA-jqcg-44mw-7w3h (critical), fixed in 2.0.8 |
python/ray/dashboard/client/package-lock.json — websocket-driver 0.7.4: GHSA-xv26-6w52-cph6 (critical), fixed in 0.7.5 |
4 vendored components with no licence
adequateThird-party code is committed into this repository without any statement of the terms it arrived under. Unknown terms are harder to deal with than inconvenient ones: an awkward licence can be complied with, an absent one cannot. Each of these needs its origin established before the code changes hands.
ci/ray_ci/deps/BUILD.bazel — 1 tracked files, no licence file |
ci/ray_ci/deps/aliases.bzl — 1 tracked files, no licence file |
thirdparty/patches — 27 tracked files, no licence file |
thirdparty/protobuf — 1 tracked files, no licence file |
6 advisories against this project in the last 24 months, 4 high or critical, 1 on CISA's known-exploited list
adequate1 of them is on CISA's known-exploited list: check first whether the release you run has the fix. The newest was published on 2026-07-01. A project that publishes advisories is telling its users about its fixes, and the count and pace say how much security work it carries. Whether 1 of them are fixed in the code you ship could not be told from this history: its version is only known to be at least 0.7.0. Answer lives in: the release you run, set against each fixed version. GHSA-q279-jhrf-cc6v (critical), GHSA-hhrp-gw25-jr43 (high), GHSA-q5fh-2hc8-f6rq (moderate) are fixed in the surveyed commit; which release carries them cannot be told from this history, which does not place the release users run. Answer lives in: the version you run. The record is incomplete: NVD: 30 records name ray with nothing tying them to this repository (CVE-2002-2036, CVE-2004-0701, CVE-2006-4049, …).
| 2025-11-26: GHSA-q279-jhrf-cc6v (critical; on CISA's Known Exploited list), fixed in 2.52.0: Critical RCE Vulnerability against Ray Devs exploitable via Browser (Safari & Firefox) due to DNS Rebinding Attack |
| 2026-07-01: GHSA-hhrp-gw25-jr43 (high), fixed in >= 2.56.0: Arbitrary code execution via ray.data.read_webdataset default decoder: pickle.loads(value) and torch.load(weights_only=False) |
| 2026-04-21: GHSA-mw35-8rx3-xf9r (high), fixed in 2.55.0: Remote Code Execution via Parquet Arrow Extension Type Deserialization |
| 2026-03-17: CVE-2026-32981 (high): A path traversal vulnerability was identified in Ray Dashboard (default port 8265) in Ray versions prior to 2.8.1. Due to improper validation and sanitization o |
| 2026-02-20: GHSA-q5fh-2hc8-f6rq (moderate), fixed in 2.54.0: Dashboard DELETE endpoints allow unauthenticated browser-triggered DoS (Serve shutdown / job deletion) |
Observability & Operations
adequateNo structured logging library is in use
adequateNothing references a structured logging library. Logs that are written to be read by a person are not searchable by a machine, which means incident response here is proportional to how well someone remembers the codebase rather than to what the tooling can find.
| searched 7,887 files for 10 structured logging libraries; 0 matched |
No error reporter is wired in
adequateNothing in the service initialises an error reporter (Sentry, Rollbar, Bugsnag, Honeybadger or an APM agent). An exception in production reaches a person only if someone is reading the logs, which usually means a customer reports it first.
| no error reporter initialised |
Sends usage data to usage-stats.ray.io by default
adequatepython/ray/_common/usage/usage_lib.py sends to usage-stats.ray.io, and python/ray/_common/usage/usage_lib.py leaves reporting on as shipped. Every installation reports to the vendor unless its operator turns this off. For a buyer running it inside their own product, that is an outbound data flow to a third party: ask what is sent, how often, and whether the buyer's customers were told.
python/ray/_common/usage/usage_lib.py:34 — reports to usage-stats.ray.io |
python/ray/_common/usage/usage_lib.py:504 — the reporting setting is on by default |
1 of 1 deployments run a single replica
adequate1 deployments set `replicas: 1`, and nothing (an autoscaler, a disruption budget) adds more. Every deploy and every node drain is downtime, and the availability the buyer inherits has a ceiling.
release/k8s_tests/ray_rayservice_template.yaml — replicas: 1 |
Fit with the Acquirer
not assessableNot rated: structurally not assessable from a repository.
Read more, or read your own
The full report on ray-project/ray: every dimension, the questions for the room, and the audit log.
Survey a repository: free on public repositories. Private code is read on your own machine by a collector whose source you can read first.