What Survived MinIO: The Post-Archive S3 Server Landscape, May 2026

On April 25, 2026, the MinIO main GitHub repository was archived as read-only.1 The community edition is no longer maintained. The path that started in December 2025 with a maintenance-mode announcement closed five months later with git push no longer accepting writes.2

This was not surprising. It was clarifying.

The pain point that named the risk before it materialized

Vendor Lock-In has been a first-class pain point in this index since launch — not as a generic concern, but as the specific risk that any single-vendor open-source project carries the latent option to relicense, paywall, or archive at the discretion of its corporate sponsor. The contract that "open-source" suggests is not enforceable; the trademark, copyright, and roadmap remain with the company.

When this index was built, MinIO was already a node. So was Vendor Lock-In. The architectural structure of mapping both in the same graph was a quiet hypothesis: that the second pain point applied to the first technology. The April 25 archive tested it.

What the archive actually does

MinIO did not delete the code. The Apache-2.0-licensed history remains forkable. What changed is that the maintained project — security patches, console, distribution pipeline, accepted pull requests — is now AIStor, the proprietary commercial offering, with AGPL v3 enforcement on remaining open code.3 Operators running community MinIO get to keep what they have, but new vulnerabilities won't be patched, the admin console is gone from the OSS distribution, Docker Hub publishing has stopped, and the AGPL viral provisions trigger if your code links against MinIO and you serve it over a network — which describes most ways MinIO is actually deployed.4

The community fork pgsty/minio restored the admin console and binary distribution pipeline.5 But forks of an upstream that aggressively guards trademarks and won't accept patches occupy a fragile position. Most teams are not relying on it as a long-term answer.

What survived: four Apache-2.0 alternatives, each with a distinct bet

Four open-source S3-compatible servers absorbed most of the migration traffic. Three of them — Alarik, RustFS, SeaweedFS — are Apache 2.0. The fourth, Garage, is AGPL v3 — meaning teams fleeing AGPL exposure on MinIO need to be careful here too. The license heterogeneity itself is information.

Server Language License Distinguishing bet 2026 status
Alarik Swift (SwiftNIO) Apache 2.0 ARC instead of GC — no pause spikes on small-object pressure; built-in Nuxt console (the feature MinIO removed) Alpha, active development
RustFS Rust Apache 2.0 Drop-in binary replacement — same disk layout, swap the executable Alpha, heavy ecosystem momentum
SeaweedFS Go Apache 2.0 O(1) disk access; default storage backend for Kubeflow Pipelines Mature, enterprise-deployed
Garage Rust AGPL v3 Masterless, geo-distributed, edge-first Production ready

The four bets are different enough that no single one inherits the MinIO crown. Alarik wins on small-object latency for telemetry, logs, and feature stores where GC jitter matters.6 RustFS wins on migration cost — the drop-in binary positioning is unique, and the community is already pushing for an NFS-mount feature that would parallel Amazon S3 Files, the AWS feature this index covered in the April 16 post.7 SeaweedFS wins on production-readiness — five years of deployment history and the Kubeflow integration mean the operational unknowns are mostly known. Garage wins on geo-distribution and edge — but the AGPL re-introduces the licensing problem people fled.

The fragmentation is the lesson. There is no longer a default. Every team has to pick by access pattern.

The format wave underneath: Parquet's successors arrive on the same schedule

While the S3 server tier was fragmenting, the file format underneath table formats was also splitting. Apache Parquet has been the default columnar format for over a decade — used by DuckDB, Trino, Spark, ClickHouse, every major lakehouse table format. It was architected for batch CPU analytics. Modern AI workloads — wide tables, sparse arrays, high-dimensional vectors, random-access RAG retrieval — exposed its structural limits, and three formats arrived to address them:

  • Lance Format — multimodal lakehouse format with native vector indices; 100× faster random access than Parquet for AI retrieval. Already established as LanceDB's native format.
  • Vortex — newly incubating at the Linux Foundation AI & Data Foundation (transitioned from SpiralDB). Zero-copy Apache Arrow integration; compute kernels run directly on encoded data via ALP for floats and FSST for strings. Shipped a DuckDB extension in January 2026.8 100× faster random access; 10–20× faster sequential scans.
  • Nimble — Meta's purpose-built format for wide-table ML feature engineering (10K+ columns), where Parquet's metadata overhead becomes prohibitive. Block encoding for bounded memory; Flatbuffers metadata for SIMD/GPU decode.

Three format successors, three different access-pattern bets. Lance for random-access AI retrieval. Vortex for general AI/analytics with engine-native compute kernels. Nimble for ultra-wide ML feature stores. None of them is a complete Parquet replacement; they are partial replacements, each addressing a workload Parquet was never designed for.

The pattern matches the server tier exactly: default-format orthodoxy gives way to access-pattern-specific selection. Parquet won when it was the only credible columnar format. With three credible successors plus Parquet's own continued role for batch BI, the era of "default to Parquet" is over.

Iceberg v3 lands the lakehouse mutation problem

The April 25 MinIO archive happened the same week that Apache Iceberg v3 consolidated as the public-preview standard across the major engines. Snowflake shipped v3 preview in March 2026; Databricks completed bidirectional Unity Catalog interop the same quarter; AWS announced support for v3 deletion vectors and row lineage in November 2025.9 Iceberg V3 brings four substantive changes:

  • Deletion Vectors — Puffin-encoded Roaring bitmaps mark logically deleted positions instead of rewriting whole Parquet files. Up to 10× faster MERGE/UPDATE on CDC streams.10
  • Row Lineage — every row carries a unique row ID and sequence-number timestamp; incremental processing pipelines can identify changed rows with zero full-table scans.
  • VARIANT type — semi-structured payloads (nested JSON, IoT telemetry, application logs) live alongside relational columns with columnar-equivalent scan performance via shredding. The Bronze→Silver→Gold normalization ETL becomes optional.
  • Native CDC detection at the spec layer.

The cost economics shifted underneath this. Amazon S3 Tables, the AWS-managed Iceberg surface, charges $0.05/GB processed plus $0.004 per 1,000 objects for managed compaction. Independent benchmarks show the same Iceberg compaction routines on self-managed EMR (m5.xlarge) cost approximately $0.0017/GB — roughly 29× cheaper at petabyte scale.11 The managed-vs-self-managed math inverts above the operational-overhead-savings break-even point. Engineering teams should benchmark their specific shape before committing to S3 Tables for the largest hot-data tables.

What this all says about openness

The MinIO archive, the Iceberg v3 finalization, the Vortex/Nimble arrival — these are not unrelated events. They are the same shape: the previous default fragmenting into multiple specialized successors, each governed differently. The MinIO replacement set fragments by license and language. The Parquet successor set fragments by access pattern. Iceberg v3 finalizes by absorbing functionality (deletion vectors, row lineage, VARIANT) that previously required external systems.

The structural lesson is the same one this index was built around: "open-source" is not a single property. License (Apache 2.0 vs AGPL), governance (foundation vs single-vendor), and trajectory (relicensing risk) are three different axes, and the post-MinIO landscape forces each of them into the open. A team that picks RustFS in May 2026 because "it's open-source like MinIO was" is repeating the mistake. The relevant questions are now: Apache 2.0 or AGPL? Foundation governance or single-vendor? Relicensing-history clean or red-flagged? This is harder than picking a default, and it is also more honest about what the choice actually involves.

The same questions apply to format selection. Lance is governed by LanceDB Inc. Vortex is now at the Linux Foundation. Nimble is at Meta. The licenses are similar; the governance is not.

The ontology delta

This wave moved the index from 222 to 229 nodes.12 The new entries are seven Technologies and Standards that did not exist (or did not fit the index's frame) two weeks ago: Alarik, Spice.ai, Hitachi Vantara, Actian VectorAI DB, Tailscale, Vortex, Nimble. Five existing nodes received substantial enrichment: MinIO (timeline reconciliation), Qdrant ($50M Series B, March 2026), ClickHouse (bidirectional Iceberg, Delta INSERT, Paimon, Langfuse acquisition), RustFS (NFS feature request, drop-in positioning), Iceberg V3 Spec (VARIANT type, GA timeline), Amazon S3 Tables (29× managed-vs-self-managed compaction economics).

No new edge verbs were needed. The accelerates and bypasses verbs added in the April 2026 wave carry this content forward without ontology pressure.

What is harder to count, but more important, is the shape the graph now reveals. Five years ago the picture would have been simple: AWS S3 + MinIO + Parquet + Iceberg. The picture in May 2026 is six S3-compatible servers, four columnar formats, three table-format specs at v3, and a dozen vendor-managed surfaces sitting on top of all of it. Default-driven architecture is over. Access-pattern-driven architecture is the only honest answer.

The pain point named the risk before the risk materialized. The graph mapped the alternatives before they had to pick up the load. That is what an index is for.

Works cited

Footnotes

  1. MinIO repository archived — public archive banner appeared April 25, 2026; CEO Harshavardhana statement: "Community has chosen not to support the project." ↩

  2. Best MinIO Alternative 2025 — Alarik — overview of the December 2025 maintenance-mode announcement and the staged withdrawal that followed. ↩

  3. MinIO Is Done With Open Source, What Are Your Options? — coverage of the licensing pivot, AIStor commercialization, and AGPL v3 enforcement. ↩

  4. MinIO AIStor product page — current commercial offering positioning AI-optimized storage as the maintained path forward. ↩

  5. pgsty/minio fork — community-maintained fork that restored the admin console and binary distribution pipeline removed from upstream. ↩

  6. Announcing Alarik on Swift Forums — design rationale for ARC over GC, Apache 2.0 licensing, and the small-object benchmark methodology vs MinIO. ↩

  7. RustFS NFS-mount feature request — community-driven push for parity with the AWS S3 Files NFS surface launched April 2026. ↩

  8. Announcing Vortex Support in DuckDB — official DuckDB extension shipped January 23, 2026; concrete integration path for engines beyond Apache Arrow. ↩

  9. The Next Era of the Open Lakehouse: Apache Iceberg v3 in Public Preview — Databricks v3 preview announcement and bidirectional Unity Catalog interop summary. ↩

  10. Complete Guide: Reading Puffin Files in Apache Iceberg — Puffin file format walkthrough including the Roaring-bitmap encoding used for deletion vectors. ↩

  11. S3 Managed Tables, Unmanaged Costs: The 20× Surprise with AWS S3 Tables — onehouse.ai benchmark comparing S3 Tables managed compaction billing against self-managed EMR Iceberg compaction. ↩

  12. LF AI & Data Foundation hosts Vortex Project — vendor-neutral governance transition for the Vortex columnar format. ↩