Fly.io has introduced the Sprites Block Device (SBD), a new filesystem replacing the modified JuiceFS system it built at launch for its Sprites product. The original setup used a rewritten SQLite metadata backend kept durable with Litestream, storing a single 100GB disk image mounted as a loop device backed by object storage. SBD keeps the same core design—object storage as the root of a Sprite's disk so it can sleep and wake across servers—but replaces the custom tooling with a simpler, more reliable implementation.
fly.io
· 2026-10-22
Infino describes how it built a retrieval engine whose full-text, vector and SQL search index lives on S3, GCS or Azure Blob Storage rather than local disk. The company organizes data into Parquet-based 'superfiles' grouped into 'supertables,' each with a manifest and term index, to reduce the number of random reads needed when querying object storage.
infino.ai
· 2026-10-15
turbopuffer, the serverless vector database used by Cursor and Notion, is overhauling its storage architecture in a project internally called 'turbopuffer v3'. Engineer Dan Harrison says the company is replacing its ANN vector index as the primary data structure, making it just one of several secondary indexes supported by a new core engine.
turbopuffer.com
· 2026-10-01
Cloudflare has released K2, a durable event streaming primitive on its Developer Platform, now in public beta. K2 stores incoming events as an ordered log built on R2 object storage, letting multiple consumers read data independently and at their own pace without losing events during downtime.
blog.cloudflare.com
· 2026-10-01
Meta has published an architectural overview of Delta, an internal object storage service designed to hold build and distribution artifacts and support disaster recovery bootstrap operations. Delta exposes just four operations—put, get, delete, and list—and prioritizes reliability and horizontal scalability over latency or storage efficiency.
engineering.fb.com
· 2026-09-21
A developer building an open-source Git server on top of object storage (Tigris) found that standard Git packfiles, designed for local disk and memory-mapped access, performed poorly when objects had to be fetched via byte-range guesses from cloud buckets. After an initial filesystem-shim approach failed at production scale, they created a new columnar packfile format built specifically for object storage, which required no changes to Git clients and performed well on real-world repositories.
tigrisdata.com
· 2026-09-16
Ampbase's founder JP explained how the company replaced a traditional relational database with Tigris, an object storage system built on FoundationDB, using only two storage primitives. To make this work, Ampbase had to manually implement unique constraints, transactions, indices, and history tables, features a normal database engine provides automatically.
tigrisdata.com
· 2026-09-08
A storage server lost power at 00:32 UTC on 17 August 2026 and stayed offline until an engineer manually restarted it at 08:22 UTC, causing an 8-hour-15-minute outage across S3-compatible storage, the container registry, and serverless and static-site deployments. The company's status page correctly flagged the incident within three minutes, but no engineer noticed it until nearly eight hours later when someone checked the page for an unrelated reason. No customer data was lost or corrupted, though failed deployments during the window require manual re-triggering.
danubedata.ro
· 2026-08-24
PicoMQ is a new messaging system that delivers durable, HTTP-accessible streams while storing data on object storage rather than dedicated disks or brokers. It lets developers create a separate stream for each use case instead of consolidating records into shared topics, with each stream independently addressable and able to scale from idle to high throughput.
picomq.com
· 2026-08-24
A blogger built a customer analytics drilldown dashboard using only a 40MB pre-aggregated Parquet 'data cube' stored in R2, an 18KB browser-based Hyparquet reader, and HTTP range requests, avoiding any database or query engine. The demo rolled up NYC's 34-million-row 311 service request dataset into a compact cube that supports fast filtering by reading only small slices of the file at a time.
hamiltonulmer.com
· 2026-08-24
GoKawiil