mirror of
https://github.com/trapexit/mergerfs.git
synced 2026-11-22 07:56:32 -05:00
Note that mergerfs itself is drive-technology agnostic, but that the SnapRAID parity and mergerfs.balance workloads commonly paired with a pool expose SMR write-amplification, so CMR is preferable for pool and especially parity members.
71 lines
3 KiB
Markdown
71 lines
3 KiB
Markdown
# Reliability and Scalability
|
|
|
|
## Is mergerfs "production ready?"
|
|
|
|
Yes.
|
|
|
|
mergerfs has been around for over a decade and used by many users on
|
|
their systems. Typically running 24/7 with constant load.
|
|
|
|
At least a few companies are believed to use mergerfs in production
|
|
environments. A number of [NAS focused operating
|
|
systems](../related_projects.md) includes mergerfs as a solution for
|
|
pooling filesystems.
|
|
|
|
Most serious issues (crashes or data corruption) have been due to
|
|
[kernel bugs](../known_issues_bugs.md#fuse-and-linux-kernel). All of
|
|
which are fixed in stable releases.
|
|
|
|
|
|
## How well does mergerfs scale?
|
|
|
|
Users have reported running mergerfs on everything from OpenWRT
|
|
routers and Raspberry Pi SBCs to multi-socket Xeon enterprise servers.
|
|
|
|
Users have pooled everything from USB thumb drives to enterprise NVME
|
|
SSDs to remote filesystems and rclone mounts.
|
|
|
|
The cost of many calls can be `O(n)` meaning adding more branches to
|
|
the pool will increase the cost of certain functions, such as reading
|
|
directories or finding files to open, but there are a number of caches
|
|
and strategies in place to limit overhead where possible.
|
|
|
|
|
|
## Are there any limits?
|
|
|
|
There is no maximum capacity beyond what is imposed by the operating
|
|
system itself. Any limit is practical rather than technical. As
|
|
explained in the question about scale mergerfs is mostly limited by
|
|
the tolerated cost of aggregating branches and the cost associated
|
|
with interacting with them. If you pool slow network filesystem then
|
|
that will naturally impact performance more than low latency SSDs.
|
|
|
|
|
|
## Should I avoid SMR (shingled) drives in a mergerfs pool?
|
|
|
|
mergerfs itself is only a proxy and has no opinion about the recording
|
|
technology of the underlying drives — on its own it neither helps nor
|
|
hurts an SMR (shingled) drive.
|
|
|
|
The caveat is the workloads mergerfs pools are commonly paired
|
|
with. SMR drives rewrite overlapping tracks and absorb bursts in a
|
|
small CMR cache zone; once that zone fills under sustained random or
|
|
large sequential writes, throughput can collapse to single-digit MB/s
|
|
until the drive reorganizes itself. Two common mergerfs patterns hit
|
|
exactly that:
|
|
|
|
* **[SnapRAID](compatibility_and_integration.md#can-i-use-mergerfs-without-snapraid-snapraid-without-mergerfs)
|
|
parity**: a full `sync` or `scrub` writes large, sustained streams
|
|
to the parity drive(s). On an SMR parity drive this can stretch a
|
|
sync from hours into days or cause timeouts.
|
|
* **Rebalancing** (e.g. `mergerfs.balance` moving files between
|
|
branches to even out used space — see [tooling](../tooling.md))
|
|
produces the same sustained write pattern on the destination drive.
|
|
|
|
For pool members, and especially for parity drives, prefer CMR
|
|
(conventional) drives. SMR is generally fine for cold,
|
|
write-once-read-many data that is not re-synced or rebalanced. Drive
|
|
data sheets rarely state the recording technology outright, so
|
|
community-maintained CMR/SMR lists and per-capacity
|
|
[price-per-terabyte comparisons](https://hddhunt.com/cheapest-hdd-per-tb/)
|
|
are useful when sizing large CMR drives for a pool.
|