<p><strong>TLDR:</strong> Jake Wharton announced that Retrofit, OkHttp, Okio, and SQLDelight are moving to the Commonhaus Foundation after Square failed to maintain stewardship following his and Jesse Wilson’s departure. AWS launched S3 Annotations, a new metadata layer supporting up to 1,000 entries and 1 GB per object, automatically indexed into Iceberg tables. Tencent Hunyuan contributed two production-hardened kernels to vLLM that cut TTFT by 24% and TPOT by 17% on 8×H20 GPUs. Cloudflare revealed that billing workloads alone account for 53% of queries on its unified data platform.</p>
<hr>
<h2>Retrofit, OkHttp, Okio, and SQLDelight move to Commonhaus Foundation after Square abandons stewardship</h2>
<p>Jake Wharton announced that Retrofit, OkHttp, Okio, and SQLDelight are relocating to a new GitHub organization under the Commonhaus Foundation at lysine.dev. The move follows Wharton and Jesse Wilson leaving Square/Cash App and finding the company did not maintain the projects in good faith. If you depend on any of these libraries — and most Android and JVM backend shops do — the governance change is worth tracking, since it affects where issues get filed, who cuts releases, and what the long-term maintenance trajectory looks like. <a href="https://daily.dev/posts/2tSon9kWP" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>AWS S3 Annotations: up to 1,000 entries and 1 GB of queryable metadata per object, GA now</h2>
<p>AWS launched S3 Annotations, which blows past the existing limits of 10 tags and 2 KB of user-defined metadata. Each object can now carry up to 1,000 annotation entries with a combined 1 GB capacity, written in JSON, XML, or YAML, and annotations can be updated independently of the object itself. They’re automatically indexed into Apache Iceberg tables via S3 Metadata, making them queryable through Athena, Redshift, or any Iceberg-compatible engine. Generally available in all regions at S3 Standard rates. <a href="https://daily.dev/posts/kjDRgvZd9" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Tencent Hunyuan contributes vLLM kernels cutting TTFT 24% and TPOT 17% on H20 GPUs</h2>
<p>Two production-hardened kernels from Tencent’s HPC-Ops library are now in vLLM’s main branch. The Attention backend replaces fixed split-KV scheduling with a per-step load-balanced dynamic scheduler, hitting up to 2.95x speedup over static scheduling on mixed-length decode batches. The MoE backend fuses the entire pipeline into a compact FP8 path, delivering 1.59x over Triton and CUTLASS at TP8/EP1. End-to-end on the Hy3 295B model across 8×H20 GPUs, the combined backends cut TTFT by ~24% and TPOT by ~17% versus vLLM defaults, with no source modifications required. <a href="https://daily.dev/posts/iVteBw42s" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Cloudflare: billing workloads drive 53% of queries on its unified data platform</h2>
<p>Cloudflare published details on its unified data platform architecture, with the headline number being that billing alone accounts for 53% of all queries processed. That’s a striking distribution — it means the platform is effectively optimized around one workload type while serving everything else. Worth reading if you’re designing multi-tenant data platforms where one workload class threatens to crowd out the rest. <a href="https://daily.dev/posts/HRqTVgtmg" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>OliveYoung cuts inventory pipeline base stock generation time 50% with Spring Batch Partitioning and Kafka fan-out:</strong> OliveYoung’s inventory squad reduced base stock generation from ~50 to ~25 minutes using Spring Batch Partitioning, decoupled data models into independent Kafka consumer groups, and cut API calls by ~90% by switching to event-based push delivery for services with their own data stores. <a href="https://daily.dev/posts/7npOO91ip" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Transactional Outbox Pattern deep-dive: SELECT FOR UPDATE SKIP LOCKED, heartbeat leases, and partial indexes in Spring Boot:</strong> A production-grade walkthrough covering concurrent workers with SKIP LOCKED, exponential backoff, heartbeat/lease columns for crash recovery, and partial indexes for performance — with guidance on when to reach for db-scheduler instead of rolling your own. <a href="https://daily.dev/posts/wMVB2ODJN" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>pg-cdc streams Postgres WAL into typed Iceberg tables on S3 with column-level access via Lake Formation:</strong> The open source tool registers entities in AWS Glue Catalog and gates reads through Lake Formation tags, giving AI agents and analysts governed, credential-free access with built-in time travel via Iceberg snapshots. <a href="https://daily.dev/posts/l9Lx1HRG2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>TimescaleDB chunk exclusion runs a 7-day range query at 222ms on 25M rows where plain PostgreSQL degrades over time:</strong> The post explains why B-tree indexes on time series tables slow down as data grows and demonstrates TimescaleDB’s chunk exclusion, continuous aggregates, and columnar compression achieving 90-98% storage reduction on older chunks. <a href="https://daily.dev/posts/lGSuvGvqw" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft, AWS, and Anthropic collectively commit billions to forward-deployed engineering programs at enterprise customer sites:</strong> Microsoft’s Frontier Company will place 6,000 engineers at customer sites backed by $2.5B; AWS launched a $1B program using 45-day sprint pods — a coordinated shift where the competitive battleground has moved from model capability to proximity to customer data and workflows. <a href="https://daily.dev/posts/J1Zhto0rp" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


Server-side development, system design, API patterns, and distributed architecture.

Backend Digest

Jake Wharton announced that Retrofit, OkHttp, Okio, and SQLDelight are moving to the Commonhaus Foundation after Square failed to maintain stewardship following his and Jesse Wilson's departure. AWS launched S3 Annotations, a new metadata layer supporting up to 1,000 entries and 1 GB per object, automatically indexed into Iceberg tables. Tencent Hunyuan contributed two production-hardened kernels to vLLM that cut TTFT by 24% and TPOT by 17% on 8×H20 GPUs. Cloudflare revealed that billing workloads alone account for 53% of queries on its unified data platform.

Retrofit and OkHttp leave Square, AWS S3 Annotations land with 1 GB metadata capacity