- [SeaweedFS on Slack](https://join.slack.com/t/seaweedfs/shared_invite/enQtMzI4MTMwMjU2MzA3LTEyYzZmZWYzOGQ3MDJlZWMzYmI0OTE4OTJiZjJjODBmMzUxNmYwODg0YjY3MTNlMjBmZDQ1NzQ5NDJhZWI2ZmY)
Also, to increase capacity, just add more volume servers by running `weed volume -dir="/some/data/dir2" -mserver="<master_host>:9333" -port=8081` locally, or on a different machine, or on thoudsands of machines. That is it!
SeaweedFS started as an Object Store to handle small files efficiently. Instead of managing all file metadata in a central master, the central master only manages volumes on volume servers, and these volume servers manage files and their metadata. This relieves concurrency pressure from the central master and spreads file metadata into volume servers, allowing faster file access (O(1), usually just one disk read operation).
SeaweedFS can transparently integrate with the cloud. With hot data on local cluster, and warm data on the cloud with O(1) access time, SeaweedFS can achieve both fast local access time and elastic cloud storage capacity. What's more, the cloud storage access API cost is minimized. Faster and Cheaper than direct cloud storage!
There is only 40 bytes of disk storage overhead for each file's metadata. It is so simple with O(1) disk reads that you are welcome to challenge the performance with your actual use cases.
On top of the object store, optional [Filer] can support directories and POSIX attributes. Filer is a separate linearly-scalable stateless server with customizable metadata stores, e.g., MySql, Postgres, Redis, Cassandra, HBase, Mongodb, Elastic Search, LevelDB, RocksDB, MemSql, TiDB, Etcd, CockroachDB, etc.
For any distributed key value stores, the large values can be offloaded to SeaweedFS. With the fast access speed and linearly scalable capacity, SeaweedFS can work as a distributed [Key-Large-Value store][KeyLargeValueStore].
Let's start one master node, and two volume nodes on port 8080 and 8081. Ideally, they should be started from different machines. We'll use localhost as an example.
The volume id is an unsigned 32-bit integer. The file key is an unsigned 64-bit integer. The file cookie is an unsigned 32-bit integer, used to prevent URL guessing.
The file key and file cookie are both coded in hex. You can store the <volumeid,filekey,filecookie> tuple in your own format, or simply store the `fid` as a string.
If stored as a string, in theory, you would need 8+1+16+8=33 bytes. A char(33) would be enough, if not more than enough, since most uses will not need 2^32 volumes.
If space is really a concern, you can store the file id in your own format. You would need one 4-byte integer for volume id, 8-byte long number for file key, and a 4-byte integer for the file cookie. So 16 bytes are more than enough.
Since (usually) there are not too many volume servers, and volumes don't move often, you can cache the results most of the time. Depending on the replication type, one volume can have multiple replica locations. Just randomly pick one location to read.
SeaweedFS applies the replication strategy at a volume level. So, when you are getting a file id, you can specify the replication strategy. For example:
When requesting a file key, an optional "dataCenter" parameter can limit the assigned volume to the specific data center. For example, this specifies that the assigned volume should be limited to 'dc1':
Usually distributed file systems split each file into chunks, a central master keeps a mapping of filenames, chunk indices to chunk handles, and also which chunks each chunk server has.
The main drawback is that the central master can't handle many small files efficiently, and since all read requests need to go through the chunk master, so it might not scale well for many concurrent users.
Instead of managing chunks, SeaweedFS manages data volumes in the master server. Each data volume is 32GB in size, and can hold a lot of files. And each storage node can have many data volumes. So the master node only needs to store the metadata about the volumes, which is a fairly small amount of data and is generally stable.
The actual file metadata is stored in each volume on volume servers. Since each volume server only manages metadata of files on its own disk, with only 16 bytes for each file, all file access can read file metadata just from memory and only needs one disk operation to actually read file data.
The architecture is fairly simple. The actual data is stored in volumes on storage nodes. One volume server can have multiple volumes, and can both support read and write access with basic authentication.
All volumes are managed by a master server. The master server contains the volume id to volume server mapping. This is fairly static information, and can be easily cached.
On each write request, the master server also generates a file key, which is a growing 64-bit unsigned integer. Since write requests are not generally as frequent as read requests, one master server should be able to handle the concurrency well.
When a client sends a write request, the master server returns (volume id, file key, file cookie, volume node url) for the file. The client then contacts the volume node and POSTs the file content.
When a client needs to read a file based on (volume id, file key, file cookie), it asks the master server by the volume id for the (volume node url, volume node public url), or retrieves this from a cache. Then the client can GET the content, or just render the URL on web pages and let browsers fetch the content.
In the current implementation, each volume can hold 32 gibibytes (32GiB or 8x2^32 bytes). This is because we align content to 8 bytes. We can easily increase this to 64GiB, or 128GiB, or more, by changing 2 lines of code, at the cost of some wasted padding space due to alignment.
All file meta information stored on an volume server is readable from memory without disk access. Each file takes just a 16-byte map entry of <64bitkey,32bitoffset,32bitsize>. Of course, each map entry has its own space cost for the map. But usually the disk space runs out before the memory does.
The local volume servers are much faster, while cloud storages have elastic capacity and are actually more cost-efficient if not accessed often (usually free to upload, but relatively costly to access). With the append-only structure and O(1) access time, SeaweedFS can take advantage of both local and cloud storage by offloading the warm data to the cloud.
Usually hot data are fresh and warm data are old. SeaweedFS puts the newly created volumes on local servers, and optionally upload the older volumes on the cloud. If the older data are accessed less often, this literally gives you unlimited capacity with limited local servers, and still fast for new data.
If the hot/warm data is split as 20/80, with 20 servers, you can achieve storage capacity of 100 servers. That's a cost saving of 80%! Or you can repurpose the 80 servers to store new data also, and get 5X storage throughput.
SeaweedFS is meant to be fast and simple, in both setup and operation. If you do not understand how it works when you reach here, we've failed! Please raise an issue with any questions or update this file with clarifications.
SeaweedFS can also store extra large files by splitting them into manageable data chunks, and store the file ids of the data chunks into a meta chunk. This is managed by "weed upload/download" tool, and the weed master or volume servers are agnostic about it.
* SeaweedFS Filer metadata store can be any well-known and proven data stores, e.g., Redis, Cassandra, HBase, Mongodb, Elastic Search, MySql, Postgres, MemSql, TiDB, CockroachDB, Etcd etc, and is easy to customized.
MooseFS chooses to neglect small file issue. From moosefs 3.0 manual, "even a small file will occupy 64KiB plus additionally 4KiB of checksums and 1KiB for the header", because it "was initially designed for keeping large amounts (like several thousands) of very big files"
Ceph can be setup similar to SeaweedFS as a key->blob store. It is much more complicated, with the need to support layers on top of it. [Here is a more detailed comparison](https://github.com/chrislusf/seaweedfs/issues/120)
SeaweedFS has a centralized master group to look up free volumes, while Ceph uses hashing and metadata servers to locate its objects. Having a centralized master makes it easy to code and manage.
Ceph uses CRUSH hashing to automatically manage the data placement, which is efficient to locate the data. But the data has to be placed according to the CRUSH algorithm. Any wrong configuration would cause data loss. SeaweedFS places data by assigning them to any writable volumes. If writes to one volume failed, just pick another volume to write. Adding more volumes are also as simple as it can be.
SeaweedFS is optimized for small files. Small files are stored as one continuous block of content, with at most 8 unused bytes between files. Small file access is O(1) disk read.
SeaweedFS Filer uses off-the-shelf stores, such as MySql, Postgres, Mongodb, Redis, Elastic Search, Cassandra, HBase, MemSql, TiDB, CockroachCB, Etcd, to manage file directories. These stores are proven, scalable, and easier to manage.
MinIO follows AWS S3 closely and is ideal for testing for S3 API. It has good UI, policies, versionings, etc. SeaweedFS is trying to catch up here. It is also possible to put MinIO as a gateway in front of SeaweedFS later.
MinIO has full-time erasure coding. SeaweedFS uses replication on hot data for faster speed and optionally applies erasure coding on warm data.
MinIO does not have POSIX-like API support.
MinIO has specific requirements on storage layout. It is not flexible to adjust capacity. In SeaweedFS, just start one volume server pointing to the master. That's all.
When testing read performance on SeaweedFS, it basically becomes a performance test of your hard drive's random read speed. Hard drives usually get 100MB/s~200MB/s.
To modify or delete small files, SSD must delete a whole block at a time, and move content in existing blocks to a new block. SSD is fast when brand new, but will get fragmented over time and you have to garbage collect, compacting blocks. SeaweedFS is friendly to SSD since it is append-only. Deletion and compaction are done on volume level in the background, not slowing reading and not causing fragmentation.
The text of this page is available for modification and reuse under the terms of the Creative Commons Attribution-Sharealike 3.0 Unported License and the GNU Free Documentation License (unversioned, with no invariant sections, front-cover texts, or back-cover texts).
