<p><strong>TLDR:</strong> Kubernetes 1.34 made Dynamic Resource Allocation generally available, replacing the old device plugin model with structured, attribute-based GPU requests. Coupang replaced database-native sequences across 100+ services with a two-tier caching architecture that handles 50,000 sequences per second for under $1,000/month. Mintlify cut a $70K/year sandbox bill to near-zero by building a virtual filesystem on top of their existing vector database. The inbox pattern, saga pattern, and hybrid RAG search all got solid implementation writeups worth bookmarking.</p>
<hr>
<h2>Kubernetes Dynamic Resource Allocation goes GA</h2>
<p>Kubernetes 1.34 promoted DRA to general availability, and it’s a meaningful change for anyone running GPU workloads. The old <code>nvidia.com/gpu: 1</code> device plugin was blunt — DRA introduces four structured objects (ResourceSlice, DeviceClass, ResourceClaimTemplate, ResourceClaim) that let you express actual hardware requirements like architecture generation and minimum VRAM. Three GPU sharing strategies are now first-class: time-slicing for sequential access, MPS for concurrent CUDA processes, and MIG for hardware-level partitioning into isolated slices on an A100. If you’re provisioning GPU nodes today, the device plugin model is the wrong abstraction to build on.</p>
<h2>Coupang replaced database sequences at 50K/sec for under $1K/month</h2>
<p>Migrating 100+ services from a relational database to DynamoDB meant replacing native sequences, which don’t exist in DynamoDB. The solution is a two-tier cache: a thick client SDK in each application handles roughly 99% of requests from local memory, a server-side cache absorbs most of the rest, and DynamoDB sees only about 0.1% of requests via conditional writes. A sliding window algorithm predicts refill timing so cache exhaustion never blocks user requests. Peak throughput is 50,000 sequences per second with 10-20 DynamoDB writes per second. The Orders team migrated 12 services in three weeks with zero downtime and fewer than 50 lines of code changed. The key insight that made this tractable: most teams didn’t actually need gap-free or globally ordered sequences.</p>
<h2>Mintlify replaced $70K/year sandboxes with a virtual filesystem</h2>
<p>The original approach for their documentation AI assistant used sandbox environments with a ~46-second p90 boot time. At scale that was going to cost over $70,000 per year. ChromaFs intercepts UNIX shell commands — grep, cat, ls, find — and translates them into Chroma vector database queries, dropping session creation to around 100ms with near-zero marginal cost. A two-phase grep uses Chroma as a coarse filter before doing exact matching, Redis caches hot results, and lazy file pointers handle large OpenAPI specs stored in S3. The design is read-only, which keeps sessions stateless and sidesteps a whole class of consistency problems.</p>
<h2>Hybrid RAG search closes the gap vector similarity alone can’t</h2>
<p>A bug report about a customer-support RAG agent returning stale policy documents is a good illustration of a real architectural gap: vector similarity has no concept of recency, tenant boundaries, or document scope. The fix is combining vector distance functions with SQL WHERE clauses, JOINs, and GROUP BY in a single query. On a 10M-row corpus, hybrid search improved Recall@5 from 72% to 94% and cut the stale document rate from 23% to under 1%, at a latency cost of 15-30ms. The post also makes a reasonable case against the vector sidecar pattern — running a separate vector database alongside your primary relational DB creates sync pipelines and consistency windows that native vector support in one database avoids entirely.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Inbox pattern implementation:</strong> Consumer-side counterpart to the Outbox pattern, using PostgreSQL <code>ON CONFLICT DO NOTHING</code> for idempotent inserts and <code>FOR UPDATE SKIP LOCKED</code> for safe concurrent batch processing — solid reference implementation covering poison messages, ordering, and monitoring.</li>
<li><strong>Nango built 200 API integrations in 15 minutes for under $20:</strong> Using OpenCode as a headless background agent across Google Calendar, Drive, HubSpot, and Slack; key finding was that agents routinely fake completion, hallucinate CLI commands, and edit test fixtures to avoid fixing bugs — verify everything with post-completion checks.</li>
<li><strong>Temporal Spring 2026 update:</strong> Worker versioning, multi-cloud replication, and Ruby SDK hit GA; Azure cloud support targeting July pre-release; Kubernetes Worker Controller with autoscaling and OpenAI Agents SDK integration are upcoming.</li>
<li><strong>Module Federation 2.0 stable:</strong> ByteDance’s Web Infra team shipped a decoupled runtime layer supporting webpack, Rspack, Rollup, Vite, and Metro, plus dynamic TypeScript type hints that auto-generate from remote modules and first-class Node.js support for SSR and BFF patterns.</li>
<li><strong>Token auth failures traced to clock skew:</strong> Sessions dying silently days after issuance were caused by client device clocks off by minutes; fix was enforcing server-side time windows exclusively and removing trust in client-reported time.</li>
<li><strong>Database selection framework:</strong> Netflix uses Cassandra for 3M writes/sec via denormalized partition-key modeling, Instagram runs PostgreSQL at billion-user scale with PgBouncer and read replicas, Twitter pre-builds timelines in Redis with fan-out-on-write — the pattern is access-pattern-first, not technology-first.</li>
<li><strong>GitOps at enterprise scale:</strong> ArgoCD alongside Jenkins and GitHub Actions, phased rollout starting with low-risk services; deployment frequency went from weekly to multiple daily deployments with rollback via Git revert.</li>
<li><strong>AI observability architecture shift:</strong> AI agents issue dozens of concurrent queries in rapid succession rather than single heavy queries, overwhelming systems built for batch reporting; the emerging answer is separating transactional and analytical workloads with a Postgres + ClickHouse split and columnar engines for high-cardinality retention.</li>
<li><strong>NetBox TurboBulk:</strong> Uses PostgreSQL’s native COPY protocol instead of ORM, hitting over 24,000 objects per second versus 100-200 with the standard REST API; a full export of tens of millions of objects took 10 minutes versus an estimated 23+ hours.</li>
<li><strong>Self-healing deployment pipeline:</strong> LangChain’s GTM Agent uses Poisson statistical testing to distinguish real regressions from noise, routes flagged errors to a triage agent that checks causal links to code diffs, then opens a PR automatically via Open SWE if confirmed.</li>
<li><strong>Fair multitenancy rate limiting:</strong> Concurrency limits are harder than rate limits in distributed systems; leaky buckets for burst smoothing, centralized limit stores with override capabilities, and simulation modes before enforcement are the practical recommendations.</li>
<li><strong>Monorepo vs polyrepo AI context tax:</strong> AI coding assistants hallucinate more in polyrepo setups due to missing cross-repository context; monorepos solve visibility but move fragmentation into CI/CD, requiring differential deployment infrastructure.</li>
<li><strong>bcachefs removed from Linux kernel:</strong> Covered in Technology Short Take 193 alongside Trivy GitHub Actions supply chain compromise, Ingress2Gateway 1.0 GA, and containerd LTS releases.</li>
</ul>


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

Backend Digest

A curated roundup of notable engineering developments: Kubernetes 1.34 promotes Dynamic Resource Allocation (DRA) to GA, replacing the device plugin model with structured GPU resource requests supporting time-slicing, MPS, and MIG. Coupang replaced database-native sequences across 100+ services with a two-tier caching architecture achieving 50K sequences/sec for under $1K/month. Mintlify eliminated a $70K/year sandbox cost by building a virtual filesystem (ChromaFs) on top of their Chroma vector database. Additional coverage includes hybrid RAG search improving Recall@5 from 72% to 94%, the inbox pattern with PostgreSQL, Module Federation 2.0, GitOps at enterprise scale, AI observability architecture, and more.

Kubernetes DRA goes GA, Coupang drops sequences for 50K/sec throughput