011 merged · 5 active
Upstream SkyPilot engineering
Production code contributed to the SkyPilot repository. One pull request is verified as merged; five more are in development, maintainer discussion, or review.
02Open-source AI infrastructure
SkyPilot · Berkeley RISELab
Contributing to SkyPilot through Berkeley RISELab: one verified pull request merged upstream, five additional PRs in active development or review, and a separate systems-research workstream on cross-cloud inference-state recovery.
- Python
- Linux CLI
- Distributed systems
- Multi-cloud
- Kubernetes
- Open source
Proof artifactResearch affiliation · UC Berkeley RISELab
Software Engineer · Berkeley RISELab
How to read this case study
Two related, separately verified workstreams
The SkyPilot engineering contribution and the Portage research project both sit within my Berkeley RISELab work, but they have different evidence and maturity levels.
02Draft · validating
Portage systems research
A separate research workstream on reusable inference state across clouds. Current evidence is analytic and simulation-based; hardware validation and the final research artifact are still in progress.
Open-source engineering
Contribution record
01 · Verified upstream contributionMerged Aug 10, 2026
PR #10286 · Redact API server credentials from CLI output
Added a centralized URL-password redaction helper and applied it at both human-readable CLI output boundaries, preventing basic-auth secrets from leaking into terminals, logs, screenshots, or agent transcripts while preserving usable endpoint context. The patch covers IPv6, malformed inputs, path/query components, and URLs without passwords.
- 3 files changed
- +104 / -4
- Maintainer approved
- Merged upstream
02 · Active contribution pipelineIn review / ongoing
5 additional SkyPilot pull requests
Five additional contributions are currently being developed, discussed with maintainers, or reviewed. They are shown as active work—not merged work—and will receive individual links and outcomes only after their public status is verified.
- 5 active PRs
- Maintainer discussions
- Not counted as merged
What needed to change
SkyPilot can accept API-server URLs containing basic-auth credentials. When those URLs are printed for humans, an unredacted password can escape into terminal history, logs, screenshots, support threads, or AI-agent transcripts—even though the credential still needs to remain intact inside authenticated code paths.
The engineering response
For merged PR #10286, added a centralized URL-password redaction helper and applied it to SkyPilot’s human-readable API-server URL output, preventing basic-auth credentials from leaking into terminals, logs, screenshots, and agent transcripts. The patch added focused tests across standard credentials, IPv6, malformed inputs, paths, queries, and no-password URLs. Five additional pull requests remain active and are not represented as merged.
The signal
Merged one verified production pull request into SkyPilot’s main branch after review, contributing a security-focused improvement to a 10K+ star platform with 14M+ downloads. Five more contributions are progressing through development, maintainer discussion, or review and will only be counted as merged when the public record confirms it.
The durable insight
Open-source infrastructure work rewards precision over novelty. Changes must preserve cross-provider contracts, survive maintainer review, account for failure states, and remain understandable to engineers who operate the system long after the original patch lands.
Engineering depth
How I engineered it
The constraints, reliability decisions, and system boundaries behind the visible product.
- 01
In merged PR #10286, introduced a centralized URL-password redaction helper at the human-readable CLI boundary so functional credentials remain available internally while terminals, logs, screenshots, and agent transcripts receive a safe endpoint.
- 02
Applied the helper to both SkyPilot API-server URL display paths and added unit coverage for standard credentials, IPv6 hosts, paths and query strings, malformed input, and URLs without passwords. The final patch changed three files with 104 additions and four deletions.
- 03
Iterated through automated and maintainer review, addressed follow-up correctness concerns, and landed the contribution upstream rather than leaving it in a fork or portfolio-only demo.
- 04
Kept the contribution boundary narrow and auditable: secrets are redacted only where the URL becomes human-readable, while the authenticated value remains available to the code paths that still require it.
Separate research workstream
NSDI-format systems research draft · Hardware validation in progressPortage: Carry, Compress, or Rebuild?
The economics of reusable inference state for agentic workflows across clouds
Portage studies a concrete recovery decision for interrupted or migrated AI agents: transfer raw KV cache, compress it first, or carry smaller application artifacts and recompute. The work combines exact transformer-state arithmetic, a closed-form cost model, discrete-event simulation, an MILP oracle, and a network transfer harness to identify when each strategy is rational.
Method
- Exact KV-cache size arithmetic across model size, context length, precision, and grouped-query attention.
- Closed-form break-even analysis linking state size, prefill throughput, target queue time, and WAN bandwidth.
- Discrete-event evaluation over 1,334 synthetic recovery events, checked against a PuLP/CBC MILP oracle.
- Token-bucket TCP harness for validating the transfer model before cross-cloud GPU measurements.
Preliminary findings
- A 32k-context Llama-3-70B request carries roughly 10.7 GB of KV state under the draft assumptions.
- On an idle target, raw KV transfer only beats recomputation above roughly 59 Gbps in the analytic model.
- With a 12-second target queue, the action frontier becomes mixed: 44.2% recompute, 23.1% raw, and 32.7% compressed.
- The proposed policy selected 92% recompute and 8% compressed transfer, matching the simulation oracle to reported precision.
Potential SkyPilot integration
- Add task-level artifact annotations so durable application state is explicit rather than inferred after failure.
- Expose SkyPilot optimizer transfer-cost estimates through a planning interface used by the recovery policy.
- Add a managed-jobs recovery hook that can rebuild from durable artifacts instead of blindly moving opaque model state.
01Configured API URL→
02Central URL parser→
03Password redaction→
04Human-readable CLI output→
05Safe logs and transcripts