Weaviate stability patches + llama.cpp warmup deprecation — June 2, 2026

🚨 Breaking

Weaviate v1.36.17 & v1.37.7 ship critical stability fixes. The release notes list no API-level breaking changes, but the patches are production-significant and should be applied immediately:

Compaction abort on cancel (lsmkv): compaction now stops immediately when a cancel signal is received (v1.36.17, v1.37.7)
Collection / shard delete path: sync-rename + async RemoveAll replaces the previous teardown sequence — faster and race-condition-safer under concurrent load

v1.36.17 is the stable branch; v1.37.7 is the current main branch. Patch whichever you track before your next peak window.

🗑️ Deprecations

llama.cpp b9471 deprecates llama_set_warmup (release). No replacement symbol is named in the release. Audit any integration that calls llama_set_warmup before upgrading to b9471 or later.

💰 Pricing

Nothing in today's brief.

🆕 New

llama.cpp shipped several builds over the last 48 hours: - b9470: Hexagon backend adds MUL_MAT, MUL_MAT_ID, FLASH_ATTN, and GDN optimizations for latest models; fixes a fused RMS-norm-mul failure path (release) - b9444: llama-server now handles If-None-Match weak ETags correctly (release) - b9442: Tokenizer support added for jina-embeddings-v2-base-zh (whitespace tokenizer, lowercase defaults to true) (release)

OpenAI Codex 0.136.0: TUI markdown links now carry OSC 8 metadata so links remain clickable in terminal; cramped tables auto-switch to key/value layout. New session archiving: /archive in TUI or codex archive / codex unarchive CLI commands; archived sessions are protected from auto-deletion (release).

🌐 AI Industry

Google published a post-mortem on using Gemini internally to build Google I/O 2026 tooling and production workflows (blog). No technical specs or model identifiers disclosed in the brief.

💡 Today's Action

If you run Weaviate and do any collection rotation under load: upgrade to v1.36.17 or v1.37.7 today. The async-delete fix directly affects data integrity during concurrent shard teardown. If you use llama.cpp, grep your codebase for llama_set_warmup now — the deprecation window starts at b9471.