--- title: Aurora PostgreSQL silently regresses CREATE DATABASE,... description: An Aurora PostgreSQL upgrade from PG14/15 to PG16 silently changed the default CREATE DATABASE strategy, turning 7-minute cloning jobs into 3-hour ordeals with... canonical: https://daily.dev/posts/aurora-postgresql-silently-regresses-create-database-dwarfstar-inference-server-gains-traction-uzxdlkjnu twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Aurora PostgreSQL silently regresses CREATE DATABASE, DwarfStar inference server gains traction | daily.dev og:description: An Aurora PostgreSQL upgrade from PG14/15 to PG16 silently changed the default CREATE DATABASE strategy, turning 7-minute cloning jobs into 3-hour ordeals with... og:url: https://daily.dev/posts/aurora-postgresql-silently-regresses-create-database-dwarfstar-inference-server-gains-traction-uzxdlkjnu og:image: https://api.daily.dev/og/posts/uzxdLKJnu.png og:image:alt: Aurora PostgreSQL silently regresses CREATE DATABASE, DwarfStar inference server gains traction og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Aurora PostgreSQL silently regresses CREATE DATABASE, DwarfStar inference server gains traction **[Backend Digest](https://daily.dev/sources/backend_digest)** · 5 min read · 0 upvotes · 0 comments ## Summary An Aurora PostgreSQL upgrade from PG14/15 to PG16 silently changed the default CREATE DATABASE strategy, turning 7-minute cloning jobs into 3-hour ordeals with no errors or warnings. Separately, Salvatore Sanfilippo's DwarfStar inference engine is quietly maturing from a personal MacBook project into something that looks like a real multi-user server. GitHub's actions/checkout v7 ships July 20 and closes the classic pwn request attack vector, but leaves several related vectors open. Netflix published a detailed writeup on their in-house LLM serving platform built on vLLM and Triton. ## Content **TLDR:** An Aurora PostgreSQL upgrade from PG14/15 to PG16 silently changed the default CREATE DATABASE strategy, turning 7-minute cloning jobs into 3-hour ordeals with no errors or warnings. Separately, Salvatore Sanfilippo's DwarfStar inference engine is quietly maturing from a personal MacBook project into something that looks like a real multi-user server. GitHub's actions/checkout v7 ships July 20 and closes the classic pwn request attack vector, but leaves several related vectors open. Netflix published a detailed writeup on their in-house LLM serving platform built on vLLM and Triton. --- ## Aurora PostgreSQL PG16 upgrade silently changes CREATE DATABASE strategy, 22-26x slowdown A database engineer managing 30+ Kubernetes clusters found that upgrading Aurora PostgreSQL from PG14/15 to PG16 caused database cloning jobs to balloon from 7 minutes to 3 hours. The culprit: PG15 quietly changed the default CREATE DATABASE strategy from FILE_COPY to WAL_LOG, and Aurora inherited it with no documentation and no warnings. WAL_LOG clones block-by-block through the write-ahead log at roughly 8 MB/s; FILE_COPY works at the filesystem level and scales with Aurora's storage throughput. The fix is a single SQL clause — STRATEGY = FILE_COPY — and benchmarks showed 22-26x speedups on databases ranging from 13 GB to 87 GB. AWS has not documented this behavioral change anywhere, which makes it a silent trap for anyone running automated upgrade pipelines. [Read more](https://daily.dev/posts/ea1iiG0ij) ## DwarfStar, Redis creator's local inference engine, hits 120 tokens/sec on 8xL40S Salvatore Sanfilippo has been shipping DwarfStar, a purpose-built C inference engine for DeepSeek V4 Flash (284B MoE), and it's starting to look like more than a hobby project. On a 128 GB M3 Max MacBook Pro it runs at around 26 tokens/sec with an 81 GB asymmetrically quantized model; on a friend's 8xL40S server with microbatching, it hit 120 tokens/sec aggregated generation and 2,000 tokens/sec prefill. The design choices are worth understanding: 2-bit quantization on MoE expert weights with higher precision kept for shared components, KV cache persisted to disk so large context windows become a storage question rather than a memory one, and distributed inference via layer splitting. The project has 15,500 GitHub stars and an MIT license. What started as a way to run a frontier model overnight on a MacBook is starting to look like a legitimate self-hosted inference server for privacy-sensitive or cost-sensitive agentic workloads. [Read more](https://daily.dev/posts/PMfzRAZGE) ## actions/checkout v7 ships July 20, closes pwn request vector but leaves others open Starting July 20, 2026, actions/checkout v7 blocks fork pull request code from being checked out inside pull_request_target and workflow_run workflows — the combination that made pwn requests possible. Workflows pinned to floating major tags like @v4 get the new behavior automatically and may break without any action on your part. What v7 does not fix is worth paying attention to: manual git/gh checkouts in run blocks, other privileged triggers like issue_comment, and any workflow using the allow-unsafe-pr-checkout escape hatch all remain vulnerable. The recommended response is to audit all privileged triggers org-wide, split trusted from untrusted workflows, and pin actions to full SHAs with Dependabot for deliberate upgrades. [Read more](https://daily.dev/posts/w87TxXU78) ## Netflix details in-house LLM serving on vLLM and Triton, including constrained decoding fix Netflix's AI Platform team published a detailed writeup on their LLM serving platform, built on vLLM and NVIDIA Triton Inference Server. The interesting part is the constrained decoding story: an initial per-request Python logits processor hit GIL-induced CPU bottlenecks under batch load, which they fixed by migrating to vLLM V1's batch-level API and reimplementing the hot path in multi-threaded C++. Other operational details include switching from TensorRT-LLM to vLLM as the primary engine, model caching on Amazon FSx to reduce cold-start latency, patching Triton's OpenAI frontend to properly pass response_format to vLLM's guided decoding, and building a unified Prometheus metrics proxy to merge vLLM and Triton observability. For teams running their own inference infrastructure, the GIL bottleneck in logits processors is a real gotcha worth knowing about before you hit it in production. [Read more](https://daily.dev/posts/yh81G6jww) --- ## Also notable - **PG Durable: Microsoft open-sources crash-proof workflow extension for PostgreSQL:** PG Durable is an open-source PostgreSQL extension from Microsoft that persists each step of a workflow to disk, providing crash-proof durable functions with built-in scheduling, automatic retries, and human-in-the-loop approval signals — no external queue required. [Read more](https://daily.dev/posts/Gzulovwc1) - **Cloudflare November 2025 outage: permissions change to ClickHouse cluster took down ~20% of the web for 90 minutes:** A permissions change to a ClickHouse cluster corrupted a bot-scoring feature file and broke Cloudflare's core proxy for roughly 90 minutes in November 2025, illustrating how the same tight coupling that makes Cloudflare effective at stopping external threats causes internal mistakes to cascade globally. [Read more](https://daily.dev/posts/yybICuX0A) - **Docker daemon runs as root since boot; rootless mode stable since 20.10 but almost nobody uses it:** Anyone in the docker group has effective root on the host via the documented API — no exploit needed — because dockerd has run as root since 2013; rootless mode has been stable since Docker 20.10 but remains opt-in and rarely adopted. [Read more](https://daily.dev/posts/KUAcoPqbw) - **DataFusion Comet Iceberg writes show 3.6x speedup for wide tables:** A new contribution to DataFusion Comet implements Iceberg writes by splitting the V2 write operator into separate write and commit steps, letting the Rust layer write Parquet files directly from Arrow batches without transposing to Spark's row format — benchmarks show a 3.6x improvement for 100-column tables end-to-end. [Read more](https://daily.dev/posts/EY9NjsbkS) - **Zig web API at 22,000 lines: strong for constrained deployments, impractical for deadline-driven work:** A developer who built a full web API backend in Zig across 22,000 lines found the ecosystem too unstable for production deadlines — libraries frequently break across Zig versions and common plumbing like auth, logging, and caching must be hand-rolled — but recommends it for IoT/ESP32 or performance-critical deployments where the 7 MB static binary size matters. [Read more](https://daily.dev/posts/IyAEmKwH7) ## Similar posts on daily.dev - [3 Hours to 7 Minutes: Fixing Aurora PostgreSQL’s Silent CREATE DATABASE Regression](https://daily.dev/posts/3-hours-to-7-minutes-fixing-aurora-postgresql-s-silent-create-database-regression-ea1iig0ij) · Medium · 0 upvotes · 0 comments --- Tags: [#security](https://daily.dev/tags/security), [#backend](https://daily.dev/tags/backend), [#postgresql](https://daily.dev/tags/postgresql), [#github-actions](https://daily.dev/tags/github-actions), [#ai-inference](https://daily.dev/tags/ai-inference) [View this post on daily.dev](https://daily.dev/posts/aurora-postgresql-silently-regresses-create-database-dwarfstar-inference-server-gains-traction-uzxdlkjnu) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/aurora-postgresql-silently-regresses-create-database-dwarfstar-inference-server-gains-traction-uzxdlkjnu","headline":"Aurora PostgreSQL silently regresses CREATE DATABASE, DwarfStar inference server gains traction","text":"An Aurora PostgreSQL upgrade from PG14/15 to PG16 silently changed the default CREATE DATABASE strategy, turning 7-minute cloning jobs into 3-hour ordeals with no errors or warnings. Separately, Salvatore Sanfilippo's DwarfStar inference engine is quietly maturing from a personal MacBook project into something that looks like a real multi-user server. GitHub's actions/checkout v7 ships July 20 and closes the classic pwn request attack vector, but leaves several related vectors open. Netflix published a detailed writeup on their in-house LLM serving platform built on vLLM and Triton.","url":"https://daily.dev/posts/aurora-postgresql-silently-regresses-create-database-dwarfstar-inference-server-gains-traction-uzxdlkjnu","datePublished":"2026-07-19T04:18:55.532Z","dateModified":"2026-07-19T04:19:22.353Z","author":{"@type":"Organization","name":"Backend Digest","logo":"https://media.daily.dev/image/upload/s--9uCSrsAb--/f_auto,q_auto/v1773839398/logos/backend_digest?_a=BAMAMiiu0","url":"https://daily.dev/sources/backend_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/backend_digest","name":"Backend Digest"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Backend Digest","item":"https://daily.dev/sources/backend_digest"},{"@type":"ListItem","position":3,"name":"Aurora PostgreSQL silently regresses CREATE DATABASE, DwarfStar inference server gains traction"}]} ```