Performance tuning — Blixt Documentation

BlixtFS’s defaults suit most deployments. When you tune, measure before and after each change with your own workload. The dashboards show where time goes.

Put the servers close to the bucket

The biggest single factor is the network path to the object store. Run BlixtFS in the same cloud region as the bucket. Cross-region and on-premises-to-cloud deployments work, but every uncached read pays the round trip.

Give it fast local disk

The chunk cache and write log live in /data. Use a dedicated SSD or NVMe volume, not the container’s root filesystem or a network disk, and make it large enough to hold the working set. A cache too small for the working set reads the same data from the bucket again and again.

Setting	Default	Effect
`cache.stop_write_percent`	95	Disk use at which new writes are refused until uploads free space
`cleaner.target_percent`	90	Disk use the cache cleaner shrinks back to
`cache.enable_compression`	true	Compress cached chunks; turn off for data that doesn’t compress, such as video

Metadata

Cold lookups and listings are answered from the database, so database latency is file-operation latency.

Setting	Default	Effect
`database.cache_size_mb`	512	In-memory metadata cache per file server
`database.preload_files`	10000	Files loaded into the cache at startup
`database.connections`	25	Database connections per server

For an external database, keep it in the same zone as the file servers.

Parallelism

Setting	Default	Effect
`general.workers`	10	Worker threads per bucket
`indexing.workers`	32	Parallel listing during indexing and consistency checks
`nfs.threads`	512	NFS server worker threads

Large buckets

Mark buckets with millions of objects lazy, so they are served before indexing finishes.
Consistency checks list the bucket. On very large buckets, lengthen indexing.fsck_sleep_seconds if checks run back to back, and rely on change notifications between them.

Scale out

When one server is saturated, a High Performance deployment spreads metadata across several file servers and data across cache and write tiers. See Scale-out.