DB Engineering Weekly: June 1–8, 2026
PostgreSQL 19 Beta 1 arrives feature-complete (SQL/PGQ graph queries, online checksum toggle, parallel autovacuum, 2× FK insert performance) with GA targeting September 2026. On the same day, Supabase closes a $500M Series F at $10B and drops Multigres v0.1 Alpha — an open-source Kubernetes Postgres operator by the original Vitess co-founder, minus sharding in this release. Weaviate v1.38.0 graduates HFresh disk-based vector indexing to GA. Redis 8.8 ships with a new array data type designed by antirez and 68–83% throughput gains on MGET/XREADGROUP. Qdrant v1.18.2 patches two security vulnerabilities including an auth whitelist bypass. Percona ClusterSync 0.9.0 unlocks MongoDB 6.0→8.0 live migration. MySQL 8.0 EOL is seven weeks out with no new Oracle tooling. SIGMOD 2026 concludes in Bengaluru (349 papers, CoddSpeed wins Best Industry Paper), and OurSQL Foundation formalizes its governance.
PostgreSQL 19 Beta 1 shipped on June 4 — feature-complete, with SQL/PGQ graph queries, online checksum toggling, parallel autovacuum, and a 2× FK insert speedup that has been in development for over a year. The same day, Supabase closed a $500M Series F at a $10B valuation and dropped Multigres v0.1 Alpha, a Kubernetes-native multi-tenant Postgres operator led by the original co-founder of Vitess. Weaviate v1.38.0 went GA on June 5, graduating HFresh disk-based vector indexing out of preview. Redis 8.8 arrived June 2 with a new array data type designed by antirez, 68–83% faster stream and MGET operations, and a built-in rate limiter. MySQL 8.0 EOL is now seven weeks away with no new migration tooling from Oracle. And SIGMOD 2026 wrapped in Bengaluru, where Microsoft Fabric's hardware-accelerated query engine won Best Industry Paper with 60+ co-authors.
PostgreSQL 19 Beta 1: five changes that affect upgrade decisions
PostgreSQL 19 Beta 1 released June 4, targeting GA in September or October 2026. 1 The Global Development Group warns that some details may change before GA, but the feature set is now locked for integration testing. Five items have concrete upgrade or evaluation implications:
SQL/PGQ graph queries. PG19 implements ISO/IEC 9075-16:2023 via a
GRAPH_TABLE expression and MATCH pattern syntax. The implementation is a rewriter, not a native graph engine: property graph definitions are metadata overlays on existing relational tables, and graph patterns compile down to relational join trees executed by the standard planner. Existing indexes apply directly, and no schema migration is required to start writing graph queries. Christophe Pettus (PGX) puts the competitive boundary clearly: for fixed-depth relationship queries — "find the accountants my CFO has emailed this quarter" — PG19 is competitive with Neo4j. For deep variable-length traversal over large graphs, Neo4j's index-free adjacency (O(degree) per hop vs. O(log n) for a B-tree join) retains an order-of-magnitude structural advantage. 2Online checksum toggle. Data checksums can now be enabled or disabled on a running cluster without a restart. PG18 flipped the default to ON; PG19 closes the 13-year gap for clusters initialized without checksums. Pettus's framing: "The thirteen-year arc runs from 'decide at initdb and live with it forever' to 'toggle it whenever you like on a live system.'" 3 One migration constraint to verify:
pg_upgrade requires matching checksum settings between source and target clusters — a mismatch aborts the upgrade.Parallel autovacuum.
autovacuum_max_parallel_workers lets a single autovacuum worker use multiple parallel workers for large table vacuums. A new scoring system re-ranks tables pending vacuum continuously, so a large table no longer blocks the queue while waiting for its assigned worker slot.2× FK insert performance. Insert workloads that trigger foreign key constraint checks see roughly 2× throughput improvement. The gain comes from reduced per-row lock overhead on the referenced table during bulk inserts — relevant for any ETL or data-loading pipeline that writes into tables with FK relationships.
Breaking changes to note. JIT compilation is now disabled by default (reverting PG17's default-on behavior).
default_toast_compression changes from pglz to lz4. RADIUS authentication has been removed entirely — any cluster using RADIUS must migrate to LDAP, GSSAPI, or certificate auth before upgrading.Trade-off callout: SQL/PGQ is worth evaluating now if you have a dedicated graph database running queries on data that also lives in Postgres. The zero-migration path — same tables, new query syntax — reduces operational footprint for shallow traversal workloads. For teams not yet on PG18, the online checksum toggle and parallel autovacuum are reasons to skip PG18 and plan directly for PG19 GA in September.
Supabase raises $500M and launches Multigres v0.1 Alpha
Supabase closed a $500M Series F at a $10B pre-money valuation on June 4, led by GIC with Stripe (returning), Georgian, and Salesforce Ventures as new investors. 4 The company reports nearly 10 million developers on the platform — double the count from eight months prior — with database starts up 600% year-over-year. CEO Paul Copplestone stated that more than 60% of new databases are now created by AI tooling rather than humans directly.
The funding announcement was paired with Multigres v0.1 Alpha, an open-source Kubernetes operator positioning itself as production-grade multi-tenant Postgres infrastructure. 5 Multigres is led by Sugu Sougoumarane, co-creator of Vitess (the sharding layer that scales MySQL at YouTube and PlanetScale), with the stated goal of bringing Vitess-grade horizontal scaling and high availability to Postgres.
The v0.1 architecture has three notable design choices:
- Consensus-based HA. Failover uses a generalized consensus protocol that can resolve split-brain scenarios without losing committed transactions — replacing the simpler primary-election approaches used by most Postgres operators.
- Dual connection pooling. A
multigateway+multipoolerstack is context-aware and eliminates the mode-selection problem that forces operators to choose between session-mode and transaction-mode pooling in PgBouncer. Each user gets an isolated pool. - Replica-side backups. pgBackRest runs against replicas, not the primary, so backup I/O doesn't contend with production writes.
The Multigres v0.1 feature notably absent: sharding. Sougoumarane's roadmap lists it as the "flagship feature," but it is not in the alpha. The operator is self-hostable on any Kubernetes cluster.
Trade-off callout: Multigres v0.1 is alpha software — not production-ready. For teams evaluating Kubernetes Postgres operators today, Percona PGO 3.0.0 (GA, last week) and Crunchy PGO v5.x remain the operationally proven options. Multigres is worth watching if horizontal sharding is a future requirement, given Sougoumarane's track record. The Supabase valuation signals sustained vendor investment in the Postgres ecosystem broadly.
Weaviate v1.38.0 GA: HFresh out of preview, Namespaces and Nested Filtering arrive
Weaviate v1.38.0 went GA on June 5, contributed by 28 developers with zero breaking changes. 6 The headline is HFresh graduating from Preview (introduced in v1.36) to Generally Available.
Loading content card…
HFresh is a disk-based vector index targeting datasets where in-memory HNSW is impractical due to memory cost. The architecture: vectors are divided into postings stored in an on-disk LSM store, queries route through a compact in-memory HNSW centroid index, and asymmetric distance computation reduces per-query memory access. This release adds nine performance improvements including normalized query vector caching, lower posting map memory usage, and retry logic for transient I/O errors. The latency tradeoff is real — HFresh operates in the hundreds-of-milliseconds range per query vs. the tens-of-milliseconds range for in-memory HNSW — but memory footprint shrinks substantially for large corpora.
Three features entered Preview status in v1.38.0:
- Namespaces — a control-plane isolation layer for multi-tenancy, built on RAFT-backed entity management with RBAC/Casbin and OIDC integration. Per-namespace object and collection limits, cascading delete, and user management across 42 PRs. API and storage format may still change before GA.
- Nested Object Filtering — allows filtering on properties inside indexed JSON objects using dotted and indexed path syntax (
cars.make,cars[0].make). Gated behindWEAVIATE_PREVIEW_NESTED_FILTERING(default off). The feature spans a 21-part implementation across LSM storage, gRPC ingress, GraphQL ingress, and position encoding. - Alter Schema Reindex — changes a property's index type at runtime without taking the collection offline. A two-phase RAFT swap barrier ensures the bucket pointer flip is atomic, eliminating the window where a BM25 query could return zero results from a stale bucket mid-reindex.
Also reaching production-ready status: Async Replication (default enabled) and Replica Movement (via a change-capture log that streams writes arriving during file copy). These two stabilize the multi-node operational story for clusters under continuous write load.
The maintenance releases — v1.37.7 and v1.36.17 (LTS) on June 2 — backport compaction abort-on-cancel, accelerated shard deletion, and async-delete idempotency to both active branches. 7
Trade-off callout: HFresh is now the correct index type for corpora where memory cost is the binding constraint and sub-100ms query latency is not required. If you are currently sharding a large HNSW dataset across multiple nodes to fit in memory, HFresh warrants a direct benchmark comparison — the operational simplification of fewer nodes may outweigh the latency regression. Namespaces and Nested Object Filtering are Preview; use them in non-production environments only.
Redis 8.8: new array type, 68–83% faster streams, built-in rate limiter
Redis 8.8 released June 2 in open source, led by Lior Kogan and featuring a new array data type designed by Salvatore Sanfilippo (antirez), Redis's original creator. 8 The release drew 228 upvotes and 107 comments on Hacker News. 9
Loading content card…
Array data type. The new type supports O(1) index access, sparse-friendly storage (empty groups cost 8 bytes), a ring buffer via
ARRING (2.2× faster than RPUSH+LTRIM), server-side aggregation with AROP (SUM, MIN, MAX, AND, OR, XOR), and pattern search via ARGREP (exact, substring, glob, and regex backed by the TRE engine). The primary target is fixed-size sliding window workloads — time-series, rate counters, and metric buffers — where list-based workarounds currently require the application to manage trimming. 10Performance benchmarks (single instance, Intel Sapphire Rapids m7i.metal-24xl): 8
| Operation | Improvement |
|---|---|
| MGET pipelined (with I/O threads) | up to 68% faster |
| XREADGROUP (COUNT 100) | up to 83% faster |
| ZADD / ZINCRBY / ZRANGEBYSCORE | up to 74% faster |
| Persistence / replication | up to 60% faster |
Window counter rate limiter. The new
INCREX command implements a window counter with bound enforcement (UBOUND), saturation support, and expiration-only-on-create semantics (ENX). This is a built-in, single-command alternative to the Lua script patterns most teams use today for sliding-window rate limiting.JSON FPHA. Explicit control over floating-point array storage format (BF16, FP16, FP32, FP64) for the JSON data type — relevant for teams using Redis JSON to cache AI embedding vectors, where precision-vs-memory tradeoffs are applied per-field.
Streams XNACK. Three explicit failure modes for stream message acknowledgment:
SILENT (transient failure, delivery counter decremented), FAIL (resource-based failure, counter unchanged), FATAL (poison message, counter set to LLONG_MAX). XNACK replaces the current pattern of inspecting PEL age and retry count to infer failure severity.Trade-off callout: The array type directly replaces
RPUSH+LTRIM sliding window patterns, with the 2.2× ring buffer speedup making the migration straightforward. The built-in rate limiter removes a common Lua dependency. If you are on Redis 7.x and running I/O thread–intensive read workloads, the 68–83% MGET/XREADGROUP gains are large enough to justify benchmarking the upgrade on your actual traffic profile before committing.Vector DB briefing: Milvus v2.6.18 and Qdrant v1.18.2
Milvus v2.6.18 (June 5) adds two schema-level features: 11
- Nullable vectors. All six vector types (
FloatVector,BinaryVector,Float16Vector,BFloat16Vector,Int8Vector,SparseFloatVector) can now be declared nullable. NULL vectors consume no additional storage, and searches skip them automatically. This matters for datasets where not all records have an embedding — previously, you needed a sentinel vector or a separate filter collection. - Element-level Struct field search. Vector searches on Struct Array fields now report the matched element's offset within the array, enabling per-element retrieval from nested data structures.
- HTTP/2 support on the Proxy REST server, with both h2c and ALPN-based TLS listeners.
The 18 bug fixes include two correctness issues: a panic on clustering keys of unsupported types, and data loss for dynamic fields after
schema evolution + partial update. Both are worth auditing if you use schema evolution against tables with dynamic fields.Qdrant v1.18.2 (June 4) is a security patch release. 12 Two vulnerabilities fixed:
- REST auth whitelist bypass (PR #9254): a specially crafted path could reach authenticated endpoints before the whitelist check, bypassing authorization entirely.
- Out-of-bounds heap read via malicious snapshot (PR #9268): an attacker-controlled snapshot with an incorrect length field could trigger a heap read beyond the allocated buffer.
Both affect Qdrant deployments with REST API exposed to untrusted networks. Upgrade immediately if you are on any v1.18.x release prior to v1.18.2.
Additional fixes: a potential infinite optimizer loop with
multi_vectors + prevent_unoptimized, MatchAny rejection of empty integer lists, and a shard snapshot streaming endpoint that previously never timed out on unresponsive clients.Trade-off callout: The Qdrant auth bypass is a high-priority patch — any instance with REST exposed without a network-layer firewall should treat this as urgent. For Milvus, nullable vector support is the most broadly applicable new feature, particularly for multi-modal pipelines where image or audio embeddings are optional fields.
Relational ecosystem: MariaDB EOL, MySQL EOL, MongoDB cross-version replication, Azure PG
MariaDB 10.6 EOL is four weeks out (July 6). 13 Three upgrade paths: switch to Enterprise Server 10.6 (extended support to August 23, 2029), upgrade to Community 11.8 (new query optimizer, built-in vector search), or migrate to MariaDB Cloud. The Enterprise path preserves current behavior but requires a support contract and forgoes the 11.8 query optimizer improvements. The June 2 Q2 corrective releases touched 11.8.8, 11.4.12, 10.11.18, and 10.6.27 — all maintenance-level fixes. 14
MySQL 8.0 EOL is seven weeks out (July 31). No new migration tooling from Oracle, Percona, or AWS appeared this week. 15 The AWS RDS upgrade guide from May 14 remains the best available reference: Blue/Green switchover under 5 seconds, three paths (in-place, Blue/Green, snapshot restore), and Extended Support pricing at approximately $73 per vCPU per month for Year 1–2 (Multi-AZ doubles the cost). MySQL 8.4 breaking changes that require code or config review:
caching_sha2_password is now the default auth plugin (mysql_native_password disabled by default), SOURCE/REPLICA terminology replaces MASTER/SLAVE in replication commands, and innodb_adaptive_hash_index defaults to OFF.Percona ClusterSync for MongoDB 0.9.0 (June 4) resolves the last major gap in MongoDB live migration tooling: cross-version replication from 6.0 to 8.0. 16 Previous versions required source and target to run the same major version. The 0.9.0 fix was simpler than expected: MongoDB 6.0 and 8.0 change-stream event formats are identical — the earlier assumption of a translation requirement was wrong. The actual blocker was DDL errors being silently swallowed, causing cascading failures. The fix surfaces DDL errors and halts the run rather than ignoring them. Full support matrix: 6.0→7.0, 6.0→8.0, 7.0→8.0, plus self-to-self for all three versions. Coverage includes both replica-set and sharded topologies. As Adnan Supic (Percona) put it: "That assumption was wrong, and being wrong about it is the most interesting part of this story." 16
Loading content card…
Azure Database for PostgreSQL Flexible Server — Christophe Pettus published the fifth installment of his managed PostgreSQL deep-dive series. 17 The structural difference from every other managed Postgres service (RDS, Aurora, Cloud SQL, AlloyDB): Azure Flexible Server HA uses PostgreSQL-level synchronous streaming replication, not block-level storage replication. Every commit on an HA-enabled primary waits for the standby's WAL flush confirmation. Zone-redundant HA adds 1–3ms per commit for the cross-AZ round trip. The blast radius scales with commit count, not commit size — so the operational mitigation is batching small writes, not reducing write volume.
One footgun documented: canceling a COMMIT that is parked on synchronous replication does not roll back the transaction. It returns a WARNING and marks the transaction as committed without quorum durability — a potential double-write path.
synchronous_standby_names is service-managed and not visible to users. The SyncRep wait event is the direct diagnostic signal for synchronous replication lag.Debezium 3.5.2.Final (June 2) is a bug-fix release: Kafka 4.1.2, CockroachDB connector default value handling, Oracle DDL parser fixes. 18 Debezium 3.6.0 remains in Beta1 (May 29).
SIGMOD 2026 and OurSQL Foundation
SIGMOD 2026 ran May 31–June 5 in Bengaluru — 349 accepted papers from 1,049 submissions (33% acceptance rate), the first time the conference has been held in India. 19
Best Research Paper: "GPM: The Gaussian Pancake Mechanism for Planting Undetectable Backdoors in Differential Privacy" (Haochen Sun and Xi He, University of Waterloo) — a privacy systems result with direct relevance to teams deploying differential privacy in data pipelines.
Best Industry Paper: "CoddSpeed: Hardware Accelerated Query Processing in Microsoft Fabric" — 60+ co-authors from Microsoft, targeting GPU-accelerated data warehousing "designed to outlive any single chip generation." 20 The paper's explicit design goal of hardware-agnostic acceleration is an engineering response to the chip-generation cycle problem in hardware-specific query offload.
Test-of-Time Award: The 2016 Snowflake Elastic Data Warehouse paper — a 10-year retrospective on the architectural decisions (storage-compute separation, virtual warehouses, multi-cluster shared data) that have since become the reference model for cloud data warehouses.
Industry-track papers with engineering relevance: CloudJump III (Alibaba — tiered storage optimization for cloud databases), ByteHouse (ByteDance — cloud-native data warehouse architecture), FuxiShuffle (Alibaba Cloud — high-throughput shuffle service), and LakeMem (disaggregated memory caching for lakehouse queries). Full texts are not yet publicly accessible.
June 5 closing keynote: Gustavo Alonso (ETH Zürich) — "A New Golden Era for Data Management."
OurSQL Foundation formalized its governance structure this week with Sunil Kamath as president, Percona, Vettabase, and Codership as founding sponsors. 21 The foundation is a 501(c)(6) non-profit with a mandate to steward community assets (public bug database, tool portal, security transparency log) independently of Oracle. It explicitly positions itself as complementary to Oracle's MySQL effort, not a fork. The founding board includes Vadim Tkachenko (Percona), Matt Lord (PlanetScale), Sunny Bains (PingCAP), and Peter Zaitsev. Whether it attracts sustained engineering contribution beyond the Percona orbit over the next 12 months is the signal to track.
Cross-engine positioning
The PostgreSQL ecosystem is consolidating around Postgres-as-platform. Supabase's $500M round — the largest single funding event in the PostgreSQL ecosystem — is primarily a bet that Postgres will absorb adjacent workloads (graph queries via SQL/PGQ, multi-tenancy via Multigres, vector search via pgvector) rather than ceding them to specialized engines. PostgreSQL 19's SQL/PGQ graph support, arriving as GA in the same quarter as Supabase's capital deployment, is not coincidental timing. For architects evaluating dedicated graph databases for shallow traversal workloads, the no-migration path into PG19 SQL/PGQ is now a concrete option to benchmark.
Weaviate is moving faster than the industry. v1.38.0 GA reaches HFresh, Async Replication, and Replica Movement all production-ready in a single release, with maintenance fixes simultaneously backported to old-LTS (v1.35), current LTS (v1.36), and stable (v1.37). The Secure MCP Server Preview in v1.38 — enabling AI agents to read and write Weaviate natively — is an early positioning move for the agentic data access layer. Qdrant's auth bypass patch this week, by contrast, is a reminder that security review in vector DB infrastructure remains uneven industry-wide.
Redis 8.8's MGET/stream gains are large enough to act on. The 68–83% throughput improvements on common operations are not micro-benchmark noise on specialized hardware. If your service is Redis-bound on read throughput or stream consumer lag, the upgrade path is straightforward — no breaking changes, no data migration.
MySQL 8.0 EOL pressure is real and Oracle is not helping. Seven weeks remain. No new tooling. Teams running MySQL 8.0 on bare metal or self-managed infrastructure should have their 8.4 upgrade plan finalized now. RDS users have Blue/Green deployment as a low-risk path. The
mysql_native_password removal is the most common application-layer breakage to test for in staging.MongoDB 6.0→8.0 live migration is now operationally feasible. Percona ClusterSync 0.9.0 closes the cross-version gap with a simpler mechanism than expected. For teams running MongoDB 6.0 who have deferred the 8.0 upgrade because of the version-matching requirement, that constraint is gone.
No independent cross-engine benchmark results this week. CMU Database Group is in semester break. The SIGMOD 2026 industry papers (CoddSpeed, CloudJump III, ByteHouse) are the closest available data, but full texts are not yet publicly accessible. ANN benchmark updates have not appeared for any vector engine in this window.
Cover: AI-generated illustration
References
- 1PostgreSQL 19 Beta 1 Released!
postgresql.org
- 2
- 3All Your GUCs in a Row: data_checksums
thebuild.com
- 4Supabase Series F
supabase.com
- 5
- 6Weaviate v1.38.0 release
github.com
- 7Weaviate GitHub releases
github.com
- 8Announcing Redis 8.8
redis.io
- 9Redis 8.8: HN discussion
news.ycombinator.com
- 10
- 11Milvus v2.6.18 release
github.com
- 12Qdrant v1.18.2 release
github.com
- 13
- 14
- 15Upgrade strategies for Amazon RDS for MySQL 8.0 to 8.4
aws.amazon.com
- 16
- 17
- 18Debezium 3.5 release series
debezium.io
- 19SIGMOD 2026 program
2026.sigmod.org
- 20From research to product: Microsoft Fabric wins best industry paper at SIGMOD 2026
community.fabric.microsoft.com
- 21
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.