v0.1.5

Changelog

What changed in Cryo, generated from CHANGELOG.md at build time. Releases are tagged on GitHub; the version in the header is the latest tag.

Updated 5 min readEdit this pageReport an issueView as Markdown
On this page

The v0.1.x tags were cut before the changelog had per-release sections, so their changes are still listed under Unreleased. Compare two tags on GitHub, for example v0.1.4...v0.1.5, to see exactly what shipped between them.

All notable changes to Cryo will be documented in this file. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased#

Added#

  • Public chart README, NOTES, and a permissive values schema. Issue forms, CodeQL, labeler, and zizmor workflows. Contributor install no longer pulls the notebook requirements.txt.
  • Open-source license metadata: LICENSE (Apache-2.0) at repo root, README license badge corrected from the TBD to Apache-2.0.
  • SECURITY.md with GitHub Security Advisory routing and supported-versions table.
  • CODE_OF_CONDUCT.md (Contributor Covenant 2.1).
  • .github/ISSUE_TEMPLATE/ (bug, feature) and .github/PULL_REQUEST_TEMPLATE.md.
  • Initial CHANGELOG.md.
  • Docs site: the header shows the latest release tag (version in site/_src/site.json), and new Changelog (rendered from this file) and Contributing pages.

Security#

  • Failed password and Google sign-ins are throttled per account (5 in 15 minutes) and per client IP (20), using Redis so the limit holds across replicas. Locked attempts get 429 with Retry-After, and the sign-in page says how long to wait. Each failure is audited as auth.login_failed (admin-only GET /api/v1/auth/audit) and counted in cryo_auth_login_failures_total. /docs, /redoc and /openapi.json are off by default in prod (CRYO_API_DOCS_ENABLED). CRYO_METRICS_BEARER_TOKEN protects the scraper /metrics. See the operator guide (#262).

Changed#

  • Docs site homepage: "Compact first, then clean up" is replaced by a scroll-driven story of one 1 TB table. It decays for 30 days, then goes through each operation in order while a pinned S3 bill shows what each step costs or saves ($110.72 to $23.53 a month in the example). Prices are S3 Standard us-east-1, and the expiry and orphan windows match the planner's defaults. The table is drawn as the brand's ice core sample and changes with each step: grains packing into blocks, old layers lifted into a tray, orphans swept into a pile. Cryo's mark narrates and fuses its cubes on optimize. Each chapter adds a plain-language analogy and a sourced cost that isn't on the bill (S3 503 throttling, request latency, Trino join order, GDPR deletes). It ends on a printed receipt and a "tables like this in your lake" estimator. A mood layer makes cost felt: the bill, its face, Cryo's mascot, the meters and the timeline rail go from green and smiling to red and crying as the bill climbs, and each step opens with a before-and-after scorecard.
  • LICENSE and deploy/helm/cryo/LICENSE are now the verbatim Apache License 2.0 text. The previous copies had reworded clauses, so GitHub could not identify the license.
  • GitHub (tech-pw/cryo) is now the only source of truth. The GitLab-to-GitHub mirror sync (scripts/sync_oss_mirror.sh, its exclude list, the .gitlab-ci.yml sync job and docs/OSS_MIRROR.md) is removed, so pull requests merged here are no longer overwritten.
  • GitHub CI installs Python dependencies from uv.lock (the same lock the image ships) instead of resolving the newest releases on every run, builds cryo-engine-go for the scanner parity tests, and skips CodeQL and the Pages deploy while the repository is private. Lint, mypy, clippy, Rust and frontend tests that had drifted are fixed.
  • New Cryo mark: small ice cubes fusing into one solid cube, replacing the faceted-iceberg mark, which was too close to the Apache Iceberg logo. Updated in docs/brand/logo/ (with dark and one-colour variants), the docs site header, homepage animation, video poster and favicon, the web app's CryoMark and favicon, and the docs-site screenshots. See docs/brand/CRYO_BRAND_GUIDE.md §6. The homepage film still shows the old mark and needs a re-render from its source.
  • README.md rewritten around images: hero banner, maintenance flow, product screenshots and architecture, each with light and dark versions. The long comparison, features and research sections are now collapsible. Images live in docs/img/ and are rebuilt by scripts/readme_images.mjs. Removed a dead link to docs/premortem.md.
  • Docs site homepage: the maintenance animation now follows one partition through five steps (scan, plan, admit, rewrite, receipt) with a live readout for each, before/after file stats, and a clickable step strip with pause and replay. Rebuilt as responsive HTML (site/_src/includes/maintenance-flow.html) instead of separate wide and compact SVGs.
  • site/scripts/preview.mjs no longer crashes on request paths starting with // or on malformed percent-escapes.
  • README.md: four explainer diagrams. What decays in an Iceberg table and which operation fixes each; how a table earns its A–F grade; how the planner chunks partitions under a GB budget, with the Trino and Spark SQL it renders; and how the governor admits a job into a free engine slot. "How it works" now walks scan → plan → admit → rewrite and record.
  • README.md: license section now links to LICENSE instead of "TBD".
  • Spark maintenance lanes run local[N,3] (task retries), cap the long-lived driver's status store, bulk-delete 1,000 S3 keys per request, retry throttled S3 calls 32 times, and may schedule on spot capacity with on-demand fallback (maintenance.laneCapacityTypes). Batch optimize rewrites only files under the scanner's small-file threshold and no longer splits files up to 10× target, orders rewrite groups files-desc, and convert_deletes removes the delete files it leaves dangling. See docs/SPARK_ENGINE_OPTIMIZATION_2026-09-25.md.

Fixed#

  • The web UI's nginx now sends Content-Security-Policy (with frame-ancestors 'none'), X-Frame-Options: DENY, X-Content-Type-Options: nosniff and Referrer-Policy on every response, error pages included. index.html and SPA routes are no-cache so a deploy shows up at once. Hashed /assets/* files are cached for a year as immutable. HSTS is a chart value (web.hsts), off by default for installs without TLS (#260).
  • Docs-vs-code drift: docs/GOVERNANCE.md previously asserted an Apache-2.0 LICENSE link whose target file did not exist.
  • stats_compute_ndv: false was ignored, so batch runs paid for a full-table compute_table_stats scan on every optimized table (31% of lane time).
  • The Spark renderer dropped every rewrite_position_deletes parameter (partial progress).
  • Spark expire_snapshots for tables with protected snapshots called a procedure Iceberg does not have (set_snapshot_ref); it now tags the snapshot with CREATE OR REPLACE TAG.
  • A batch lane whose pod was deleted (eviction, drain, spot reclaim) was never noticed and stranded its tables until the run deadline.
  • Maintenance runs stayed queued forever after all their jobs finished: the run finalizer only looked at running runs, and maintenance runs are created queued. Finalizing also no longer drops the run's actor, requested operations and correlation_id from its summary.

0.0.0 — Initial public release#

Pre-release history predates this changelog. See git log for the full commit history.

Esc

↑↓ to navigate↵ to open