InfoQ Homepage Distributed Systems Content on InfoQ
-
Uber Builds MCP Gateway to Govern Access to 800+ Servers and 5,000+ AI Agent Tools
Uber has built an MCP Gateway to standardize how AI agents discover and invoke internal services across more than 800 MCP servers and 5,000 tools. The gateway separates tool registration from runtime execution and includes automated API discovery, access controls, sensitive data redaction, incremental tool discovery through Omni MCP, and response filtering.
-
Netflix Treats GPU Reliability as a Distributed Systems Problem
Netflix has detailed how it is building a reliability platform for large-scale GPU training and inference, using layered health checks, automated remediation, and fault-tolerant training to keep jobs running despite hardware failures.
-
How Anthropic Redesigned Source Control for AI-Driven Development
Anthropic recently detailed how it migrated its monorepo from Git to Scion. Scion is a source control system that moves repository operations to the server side and uses content-addressed storage and tiered caches for agent and CI workloads. The design sacrifices offline independence for less local data and shared infrastructure.
-
CSIRO Uses Serverless Architecture to Query 100x More Genomic Variants on AWS
CSIRO has developed Serverless Beacon, a serverless architecture for querying genomic variants without copying VCF files into a central database. The system combines Amazon S3, Lambda, Athena and DynamoDB, using metadata indexes and HTTP byte range requests to query compressed genomic regions. Tests reported 1.52 second average latency for 10,000 base pair queries.
-
Netflix Bridges AWS IAM and Internal Identity with Workload Attestation
Netflix has developed a workload attestation mechanism for Apache Spark jobs running on Amazon EMR. The design combines AWS IAM identity with signed claims from Netflix’s control plane before issuing short-lived internal X.509 certificates, allowing managed compute workloads to authenticate to internal services without directly injecting credentials.
-
Pinterest Uses Event Time Watermarks to Finalize Petabyte Scale Data Partitions
Pinterest has introduced event time-based partition finalization for its database ingestion framework. The mechanism uses Apache Flink to track event time and Apache Iceberg metadata to publish a non-regressing watermark, helping downstream jobs determine when continuously ingested CDC data is sufficiently complete for processing.
-
Grab Redesigns Counter Service Storage for 50% Lower P99 Latency
Grab migrated its high-volume Counter Service from a wide column database to Aerospike using storage abstraction, shadow traffic, data parity validation, and gradual traffic migration. The redesigned data model consolidated time buckets into map-based records. Grab reports about 50% lower production p99 read latency, 1 TB versus 3 TB of disk usage, and 45% to 50% lower cost per node.
-
Cursor Uses S3 WAL to Scale Git Storage to More than 300 Pushes per Second
Cursor has introduced Continuity, a Git storage architecture that uses an S3 backed write ahead log as the source of truth. The design turns local NVMe repositories into warm caches and separates replica coordination from consistency. Cursor reports linear read scaling with up to 100 replicas and more than 300 pushes per second with S3 Express One Zone in synthetic tests.
-
Meta’s ZGateway Cuts ZippyDB Connections 19x While Handling 1B+ Operations per Second
Meta has introduced ZGateway, a stateless proxy for ZippyDB that centralizes connection management, traffic routing, caching, load balancing, and admission control. The gateway handles more than 1 billion operations per second and about 40% of ZippyDB traffic, while Meta’s model estimates a 19x reduction in persistent connections.
-
Stateless MCP Removes Session Affinity Requirements for AWS Server Deployments
AWS details how the latest Model Context Protocol specification removes protocol-level sessions, sticky-session requirements, and session storage for remote MCP servers. The change enables independent request routing and simpler horizontal scaling while shifting application state, retries, observability, and idempotency concerns to other layers.
-
Cloudflare Cuts 100 TB of Memory from 1.1.1.1 DNS Cache
Cloudflare redesigned the in-memory representation of its Big Pineapple DNS cache, reducing the per-entry footprint by 56% and freeing roughly 100 TB of working-set memory across its fleet. The Rust-based changes also increased cache insertion throughput by 43% and reduced lookup latency by 19%, while enabling Cloudflare to increase cache capacity without additional memory.
-
Uber Redesigns M3DB Sharding with Subclusters to Limit Failure Impact
Uber has redesigned shard placement in M3DB with fixed size subclusters to limit the impact of node failures, maintenance, and cluster scaling. The approach bounds shard dependencies, preserves replica isolation, and uses a greedy algorithm to select shard migrations while avoiding a separate rebalancing pass and unnecessary data movement.
-
Dropbox Evolves Riviera Content Processing Platform to Support AI Workloads
Dropbox has evolved Riviera from a file preview service into a universal content processing platform supporting more than 300 file formats and over 100 transformation capabilities. Processing hundreds of thousands of transformations per second, Riviera now supports Search, Replay, Sign, and Dash, while its APIs enable asynchronous content extraction for AI and RAG workflows.
-
Agoda Replaces 72-Shard SQL Server Price Cache with DragonflyDB
Agoda migrated its 1.5 TB hotel Price Cache from 72 SQL Server shards to DragonflyDB to handle growing read and write volumes. The migration used staged dual reads, parity validation, gradual traffic shifting, and decentralized failover detection. Agoda reports an approximately eightfold reduction in P99 read latency, with two DragonflyDB clusters providing high availability.
-
Netflix Reworks Conductor for 420 Million Monthly Workflow Executions and 10X Larger Workflows
Netflix has reworked its Conductor workflow orchestration engine to handle larger workloads, increasing supported workflow size from about 2,500 to 30,000 tasks and reducing p99 workflow evaluation latency by about 40%. Conductor 4.0 separates workflow metadata from task data, moves evaluation to asynchronous processing, and introduces dynamic worker allocation and concurrency controls.