#object #read #latency #constant #random #scale #blob

app blobd

Blob storage designed for huge amounts of random reads and small objects with constant latency

4 releases (2 breaking)

0.2.1 Apr 13, 2023
0.2.0 Apr 7, 2023
0.1.0 Apr 6, 2023
0.0.1 Mar 28, 2023

#842 in Concurrency

SSPL-1.0

105KB
2.5K SLoC

blobd

  • Scales to millions of random concurrent partial reads over trillions of objects (tiny or huge) at constant disk-level latency.
  • Asynchronous replication and event streaming.
  • Batch creation API for very high transfer and creation rates with many small objects.
  • HTTP RESTful API with support for CORS, range requests, presigned URLs, and HTTP/2.

Design

  • On-device configurable fixed-size hash map with linked list of objects on the heap. The entire device is mapped to memory.
  • Optimised for reads, then creates, then deletes. There is no way to list objects.
  • Create an object, then concurrently write its data in 16 MiB parts, then commit it.
  • Objects are immutable once committed. Versioning is currently not possible. An object replaces all other objects with the same key when it's committed (not created).
  • Only the size is stored with an object. No other metadata is collected, and custom metadata cannot be set.
  • The device must be under 256 TiB. Objects are limited to 1 TiB. The peak optimal amount of objects stored is around 140 trillion.
  • Uncommitted objects may be deleted after 7 days. Space used by objects may not be immediately freed when the object is deleted.

History

This project used to be called Turbostore and was written entirely in C; you can still see the code here.

Dependencies

~15–24MB
~339K SLoC