JuiceFS performs 175-227x slower than ZeroFS

[solracsf]

https://www.zerofs.net/zerofs-vs-juicefs

ZeroFS demonstrated 175-227x higher throughput for write operations while using only S3 for storage. JuiceFS, which requires both S3 and a separate metadata database, experienced high failure rates and significantly slower performance across all tests.

The tests also revealed differences in resource consumption: JuiceFS used 31x more storage space and generated up to 293x more S3 API calls for the same workload, which would impact operational costs in production environments.

Why such a HUGE difference? 🤔

[jiefenghuang]

ZeroFS by default write data to SlateDB first then uploads data asynchronously to S3, which is similar to how JuiceFS uses BadgerDB or a local Redis as metadata engine, plus writeback enabled to achieve a similar effect.
ZeroFS is a single-node architecture, while JuiceFS is more focused on a distributed file system that can scale horizontally.

[solracsf]

Sure. But how to explain also this?

ZeroFS

• Consistent sub-millisecond latencies for file operations
• 100% success rate across all benchmarks
• Completed all real-world tests

JuiceFS

• Failed 92% of data modification operations
• Failed 97% of append operations
• Unable to complete Rust compilation after 69 minutes
• Errors: "No such file or directory (os error 2)" on file operations

[jiefenghuang]

First of all, one issue is that the documentation does not explain how JuiceFS was configured (very likely default configuration) or deployed, so it’s impossible to fully reproduce the setup — in particular, we don’t know which metadata engine was used.
As I mentioned above, this is not like-for-like comparison test, and itself doesn’t make sense.

As for these errors, based on the results, the object storage performance seems rather poor, and it might have hit its limit, causing write failures.
We probably won’t spend time on this — our own test cases already cover many aspects, including POSIX and benchmarks. If you find any problems, feel free to open an issue.