Skip to main content

​
SQLite

  • Group reads into one transaction. In autocommit mode, every query makes two file-lock calls and a stat of the database file. Wrap a group of reads in one transaction, or set PRAGMA locking_mode=EXCLUSIVE when a single process owns the database. Inside one transaction, point reads on Archil took 3 to 4 µs each, the same as on a local disk.
  • Batch writes into transactions. Inserting 1M rows in one transaction took 11 s, while 10,000 single-row transactions took 24 s with synchronous=FULL.
synchronous=NORMAL ran about 10x faster than FULL on Archil. In WAL mode, NORMAL is SQLite’s own recommendation for most applications: the database stays consistent, but the last transactions committed before a machine crash can roll back. Keep FULL if every committed transaction must survive a crash.

​
Prefer large sequential I/O

Large sequential I/O is a strength on Archil: cold sequential reads ran at about 1.2 GB/s from a single client. Where you can, prefer fewer, larger files and large read and write sizes.

​
Dependency installs

Installing dependencies directly on a disk works well: npm ci installed 598 packages in 7.6 s. Like a local file system, recently written files are staged on the client until an fsync or a background flush. When the installed files must be durable before you continue, run archil sync on the mount point. See Ensuring that data is persisted on shutdown.

​
Need help with your workload?

The Archil team works directly with application teams to tune performance for specific workloads. Email support@archil.com and we’ll help you get the most out of your disk.
For the full benchmark against Amazon EFS, see Solution Comparison.