search

yandex-cloud / geesefs

Finally, a good FUSE FS implementation over S3
Other
691 stars 45 forks source link

readme

GeeseFS is a high-performance, POSIX-ish S3 (Yandex, Amazon) file system written in Go

Overview

GeeseFS allows you to mount an S3 bucket as a file system.

FUSE file systems based on S3 typically have performance problems, especially with small files and metadata operations.

GeeseFS attempts to solve these problems by using aggressive parallelism and asynchrony.

Also check out our CSI S3 driver (GeeseFS-based): https://github.com/yandex-cloud/csi-s3

POSIX Compatibility Matrix

GeeseFS rclone Goofys S3FS gcsfuse
Read after write + + - + +
Partial writes + + - + +
Truncate + - - + +
fallocate + - - - -
chmod/chown Y - - + -
fsync + - - + +
Symlinks Y - - + +
Socket files Y - - + -
Device files Y - - - -
Custom mtime Y + - + +
xattr + - + + -
Directory renames + + * + +
readdir & changes + + - + +

Y Only works correctly with Yandex S3.

* Directory renames are allowed in Goofys for directories with no more than 1000 entries and the limit is hardcoded.

List of non-POSIX behaviors/limitations for GeeseFS:

In addition to the items above:

Stability

GeeseFS is stable enough to pass most of xfstests which are applicable, including dirstress/fsstress stress-tests (generic/007, generic/011, generic/013).

See also Common Issues.

Performance Features

GeeseFS rclone Goofys S3FS gcsfuse
Parallel readahead + - + + -
Parallel multipart uploads + - + + -
No readahead on random read + - + - +
Server-side copy on append + - - * +
Server-side copy on update + - - * -
Partial object updates +* - - - -
xattrs without extra RTT +* - - - +
Dir preload on file lookup + - - - -
Fast recursive listings + - * - +
Asynchronous write + + - - -
Asynchronous delete + - - - -
Asynchronous rename + - - - -
Disk cache for reads + * - + +
Disk cache for writes + * - + -

* Recursive listing optimisation in Goofys is buggy and may skip files under certain conditions

* S3FS uses server-side copy, but it still downloads the whole file to update it. And it's buggy too :-)

* rclone mount has VFS cache, but it can only cache whole files. And it's also buggy - it often hangs on write.

* xattrs without extra RTT only work with Yandex S3 (--list-type=ext-v1).

* Partial object updates (PATCH) only work with Yandex S3.

Partial object updates (PATCH)

With Yandex S3 it is possible to do partial object updates (data only) without server-side copy or reupload.

Currently the feature can be enabled with --enable-patch and will be enabled by default for YC S3 in the future.

Enabling patch uploads has the following benefits:

Note: new files, metadata changes and renames are still flushed to S3 as multipart uploads.

Installation

$ git clone https://github.com/yandex-cloud/geesefs
$ cd geesefs
$ go build

Usage

$ cat ~/.aws/credentials
[default]
aws_access_key_id = AKID1234567890
aws_secret_access_key = MY-SECRET-KEY
$ $GOPATH/bin/geesefs <bucket> <mountpoint>
$ $GOPATH/bin/geesefs [--endpoint https://...] <bucket:prefix> <mountpoint> # if you only want to mount objects under a prefix

You can also supply credentials via the AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY environment variables.

To mount an S3 bucket on startup make sure the credential is configured for root and add this to /etc/fstab:

bucket    /mnt/mountpoint    fuse.geesefs    _netdev,allow_other,--file-mode=0666,--dir-mode=0777    0    0

You can also use a different path to the credentials file by adding ,--shared-config=/path/to/credentials.

See also: Instruction for Azure Blob Storage.

Windows

Everything is the same after installing WinFSP and GeeseFS, except that GeeseFS can't daemonize so you have to create a system service manually if you want to hide the console window.

You can put credentials to C:\Users\<USERNAME>\.aws\credentials, or you can put credentials in any file and specify it with --shared-config file.txt, or you can use environment variables:

set AWS_ACCESS_KEY_ID=...
set AWS_SECRET_ACCESS_KEY=...

And then start GeeseFS with geesefs <bucket> <mountpoint>, where <mountpoint> is either a drive (like K:) or a non-existing directory. Example:

geesefs-win-x64.exe testbucket K:

Benchmarks

See bench/README.md.

Configuration

There's a lot of tuning you can do. Consult geesefs -h to view the list of options.

Common Issues

Memory Limit

New since 0.37.0: metadata cache memory usage is now also limited, OOM errors caused by metadata should now go away.

GeeseFS uses RAM for two purposes:

  1. Metadata (file listings). One metadata entry uses ~1 KB of data. Total number of cached entries is limited by --entry-limit and --stat-cache-ttl, because non-expired entries can't be removed from cache. Modified entries and entries with open file/directory descriptors are also never removed from cache. Cache TTL is 60 seconds by default, cached entry limit is 100000 by default, but in reality GeeseFS is able to list files faster, so the actual limit will be ~250000 when doing a plain listing of a very long bucket.

  2. Data. Default data cache limit in GeeseFS is 1 GB (--memory-limit). GeeseFS uses cache for both read buffers when it needs to load data from the server and write buffers when user applications write data.

    At the same time, default "large" readahead setting is set to 100 MB which is optimal for linear read performance.

    However, that means that more than 10 processes trying to read large files at the same time may exceed the memory limit by requesting more than 1000 MB of buffers and in that case GeeseFS will return ENOMEM errors to some of them.

    You can overcome this problem by either raising --memory-limit (for example to 4 GB) or lowering --read-ahead-large (for example to 20 MB).

Maximizing Throughput

If you have a lot of free network bandwidth and you want to achieve more MB/s of linear write speed, make sure you're writing into multiple files (not just 1) and start geesefs with the following options:

geesefs --no-checksum --memory-limit 4000 \
    --max-flushers 32 --max-parallel-parts 32 --part-sizes 25

This increases parallelism at cost of reducing maximum file size to 250 GB (10000 parts * 25 MB) and increasing memory usage. With a lot of available network bandwidth you'll be able to reach ~1.6 GB/s write speed. For example, with fio:

fio -name=test -ioengine=libaio -direct=1 -bs=4M -iodepth=1 -fallocate=none \
    -numjobs=8 -group_reporting -rw=write -size=10G

Concurrent Updates

GeeseFS doesn't support concurrent updates of the same file from multiple hosts. If you try to do that you should guarantee that one host calls fsync() on the modified file and then waits for at least --stat-cache-ttl (1 minute by default) before allowing other hosts to start updating the file. Other way is to refresh file/directory cache forcibly using setfattr -n .invalidate <filename>. This forces GeeseFS to recheck file/directory state from the server. If you don't do that you may encounter lost updates (conflicts) which are reported in the log in the following way:

main.WARNING File xxx/yyy is deleted or resized remotely, discarding local changes

Concurrent PATCH

When using Yandex S3, it is possible to concurrently update a single object/file from multiple hosts using PATCH method (--enable-patch). However, concurrent changes are not reported back to the clients, so you have to stop all writes and refresh the inode cache to see the changes made by other hosts.

It is strongly advised to write data in non-overlapping ranges to avoid conflicts. If you do overlapping writes then you have to guarantee that your hosts coordinate and serialize their updates by doing an fsync() before the next overlapping write.

Also note that even with PATCH, best performance is achieved when writes are aligned with object part boundaries (i.e. with 5 MB chunks for the first 5 GB by default), because the server may do internal read-modify-write for patched object parts, and it also may return an error if too many parallel requests try to modify the same object part. Such "PATCH failures" are reported in the log in the following way:

main.WARNING Failed to patch range %d-%d of file %s (inode %d) due to concurrent updates

Normally, GeeseFS retries requests failed in this way, but if you don't want it to retry them, you can also choose to drop the cached update by enabling --drop-patch-conflicts.

Asynchronous Write Errors

GeeseFS buffers updates in memory (or disk cache, if enabled) and flushes them asynchronously, so writers don't get an error from an unsuccessful write. When an error occurs, GeeseFS keeps modifications in the cache and retries to flush them to the server later. GeeseFS tries to flush the data forever until success or until you stop GeeseFS mount process. If there is too much changed data and memory limit is reached during write, write request hangs until some data is flushed to the server to make it possible to free some memory.

fsync

If you want to make sure that your changes are actually persisted to the server you have to call fsync on a file or directory. Calling fsync on a directory makes GeeseFS flush all changes inside it. It's stricter than Linux and POSIX behaviour where fsync-ing a directory only flushes directory entries (i.e., renamed files) in it.

If a server or network error occurs during fsync, the caller receives an error code.

Example of calling fsync. Note that both directories and files should be opened as files:

#!/usr/bin/python

import sys, os
os.fsync(os.open(sys.argv[1], os.O_RDONLY))

Command-line sync utility and syncfs syscall don't work with GeeseFS because they aren't wired up in FUSE at all.

Troubleshooting

If you experience any problems with GeeseFS - if it crashes, hangs or does something else nasty:

License

Licensed under the Apache License, Version 2.0

See LICENSE and AUTHORS

Compatibility with S3

geesefs works with:

It should also work with any other S3 that implements multipart uploads and multipart server-side copy (UploadPartCopy).

Services known to be broken:

Important note: you should mount geesefs with --list-type 2 or --list-type 1 options if you use it with non-Yandex S3.

The following backends are inherited from Goofys code and still exist, but are broken:

References