Skip to content

Latest commit

 

History

History
278 lines (217 loc) · 10.3 KB

README.md

File metadata and controls

278 lines (217 loc) · 10.3 KB
Raw

0-stor_v2

zstor is an object encoding storage system. It can be run in either a daemon - client setup, or it can perform single actions without an associated daemon, which is mainly useful for uploading/retrieving single items. The daemon is part of the same binary, and will run other useful features, such as a repair queue which periodically verifies the integrity of objects.

Storage and data integrity

Zstor uses 0-db's to store the data. It does so by splitting up the data in chunks and distributing them over N 0-db's.

C4Component
title Zstor setup

Component(zstor, "Zstor instance")

Deployment_Node(zerodbgroup0,"0-db group", ""){
    System(zerodb1,"0-db 1")
    System(zerodb2,"0-db 2")
}
Deployment_Node(zerodbgroup1,"0-db group", ""){
    System(zerodbx,"0-db ...") 
    System(zerodbn,"0-db N") 
}

Rel(zstor, zerodb1, "")
Rel(zstor, zerodb2, "")
Rel(zstor, zerodbx, "")
Rel(zstor, zerodbn, "")
Loading

Zstor uses forward looking error correcting codes (FLECC) for data consistency and to protect against data loss.

This means zstor constantly tries to spread the data over N, being the expected_shards, 0-db's.

As long as there are M (minimal_shards), M being smaller than N off course, chunks of data intact, zstor can recover the data.

Expected setup

Currently, zstor expects a stable system to start from, which is user provided:

  • zstor has a redundancy configuration which introduces the notion of groups: a group is a list of 0-db's which have an inherent larger risk of going down together. For example, grid 0-db's which are deployed on the same farm.

Daemon - client usage vs standalone usage

The daemon, or monitor, can be started by invoking zstor with the monitor subcommand. This starts a long running process, and opens up a unix socket on the path specified in the config. Regular command invocations (example "store") of zstor will then read the path to the unix socket from the config, connect to it, send the command, and wait until the monitor daemon returns a response after executing the command. This setup is recommended as:

  • It exposes optional metrics for prometheus to scrape.
  • Only a single upload/download of a file happens at once, meaning you won't burn out your whole cpu by sending multiple upload commands in quick succession.

If the socket path is not specified, zstor will fall back to its single command flow, where it executes the command in process, and then exits. Invoking zstor multiple times in quick succession might cause multiple uploads to be performed at the same time, causing multiple cpu cores to be used for the encryption/compression.

Current features

Supported commands

  • Store data in multiple chunks on zdb backends, according to a given policy
  • Retrieve said data, using just the path and the metadata store. Zdbs can be removed, as long as sufficient are left to recover the data.
  • Rebuild the data, loading existing data (as long as sufficient zdbs are left), reencoding it, and storing it in (new) zdbs according to the current config
  • Check a file, returning a 16 byte blake2b checksum (in hex) if it is present in the backend (by fetching it from the metastore).
  • Status : get statistics about backends

Other features

  • Config file hot reloading. If the config file is editted, SIGUSR1 can be send to a running 0-stor process to reload the configuration. Only the backends in the new configuration are used. Data is not rebuild on these new backends as long as the old backends are still operational
  • Monitoring of active 0-db backends. An active backend is considered a backend that is tracked in the config, which has sufficient space left to write new blocks.
  • Repair queue: periodically, all 0-db's used are checked, to see if the are still online. If a 0-db is unreachable, all objects which have a chunk stored on that 0-db will be rebuild on fully healthy 0-db's.
  • Prometheus metrics. The metrics server is bound to all interfaces, on the port specified in the config. The path is /metrics. If no port is set in the config, the metrics server won't be enabled.

Building

Make sure you have the latest Rust stable installed. Clone the repository:

git clone https://github.com/threefoldtech/0-stor_v2
cd 0-stor_v2

Then build with the standard toolchain through cargo:

cargo build

This will produce the executable in ./target/debug/zstor_v2.

Static binary

On linux, a fully static binary can be compiled by using the x86_64-unknown-linux-musl target, as follows:

cargo build --target x86_64-unknown-linux-musl --release

Config file

Running zstor requires a config file. An example config, and explanation of the parameters is found below.

Example config file

minimal_shards = 10
expected_shards = 15
redundant_groups = 1
redundant_nodes = 1
root = "/virtualroot"
socket = "/tmp/zstor.sock"
prometheus_port = 9100
zdb_data_dir_path = "/tmp/0-db/data"
max_zdb_data_dir_size = 25600

[encryption]
algorithm = "AES"
key = "0000000000000000000000000000000000000000000000000000000000000000"

[compression]
algorithm = "snappy"

[meta]
type = "zdb"

[meta.config]
prefix = "someprefix"

[meta.config.encryption]
algorithm = "AES"
key = "0101010101010101010101010101010101010101010101010101010101010101"

[[meta.config.backends]]
address = "[2a02:1802:5e::dead:beef]:9900"
namespace = "test2"
password = "supersecretpass"

[[meta.config.backends]]
address = "[2a02:1802:5e::dead:beef]:9901"
namespace = "test2"
password = "supersecretpass"

[[meta.config.backends]]
address = "[2a02:1802:5e::dead:beef]:9902"
namespace = "test2"
password = "supersecretpass"

[[meta.config.backends]]
address = "[2a02:1802:5e::dead:beef]:9903"
namespace = "test2"
password = "supersecretpass"

[[groups]]
[[groups.backends]]
address = "[fe80::1]:9900"

[[groups.backends]]
address = "[fe80::1]:9900"
namespace = "test"

[[groups]]
[[groups.backends]]
address = "[2a02:1802:5e::dead:babe]:9900"

[[groups.backends]]
address = "[2a02:1802:5e::dead:beef]:9900"
namespace = "test2"
password = "supersecretpass"

Config file explanation

  • minimal_shards: The minimum amount of shards which are needed to recover the original data.
  • expected_shards: The amount of shards which are generated when the data is encoded. Essentially, this is the amount of shards which is needed to be able to recover the data, and some disposable shards which could be lost. The amount of disposable shards can be calculated as expected_shards - minimal_shards.
  • redundant_groups: The amount of groups which one should be able to loose while still being able to recover the original data.
  • redundant_nodes: The amount of nodes that can be lost in every group while still being able to recover the original data.
  • root: virtual root on the filesystem to use, this path will be removed from all files saved. If a file path is loaded, the path will be interpreted as relative to this directory
  • socket: Optional path to a unix socket. This socket is required in case zstor needs to run in daemon mode. If this is present, zstor invocations will first try to connect to the socket. If it is not found, the command is run in-process, else it is encoded and send to the socket so the daemon can process it.
  • zdb_data_dir_path: Optional path to the local 0-db data file directory. If set, it will be monitored and kept within the size limits. This is primarily used when 0-stor is running as part of a QSFS deployment. In this case, a 0-db-fs instance is running, which is using a local 0-db as read/write cache. When this option is set, the size of this cache is monitored, and if needed the least recently accessed files are removed.
  • max_zdb_data_dir_size: Maximum size of the data dir in MiB, if this is set and the sum of the file sizes in the data dir gets higher than this value, the least used, already encoded file will be removed.
  • zdbfs_mountpoint: Optional path of a 0-db-fs mount. If present, a syscall will be executed periodically to retrieve file system statistics, which will then be exposed through the build-in prometheus server.
  • prometheus_port: An optional port on which prometheus metrics will be exposed. If this is not set, the metrics will not get exposed.
  • encryption: configuration to use for the encryption stage. Currently only AES is supported. The encryption key is 32 random bytes in hexadecimal form.
  • compression: configuration to use for the compression stage. Currently only snappy is supported
  • meta: configuration for the metadata store to use, currently only zdb is supported
  • groups: The backend groups to write the data to.

Explanation:

Metadata

When data is encoded, metadata is generated to later retrieve this data. The metadata is stored in 4 0-dbs, with a given prefix.

For every file, we get the full path of the file on the system, generate a 16 byte blake2b hash, and hex encode the bytes. We then append this to the prefix to generate the final key.

The key structure is: /{prefix}/meta/{hashed_path_hex}

The metadata itself is encrypted, binary encoded, and then dispersed in the metadata 0-dbs.

Metadata cluster requirements

Since the metadata is also encoded before being stored, we need to know the used encoding to be able to decode again. Since we can't store metadata about metadata itself, this is a static setup. As said, at present there are 4 metadata storage 0-db's defined. Since the key is defined by the system, these must be run in user mode. At the moment, it is not possible to define more metadata stores as can be done with regular data stores.

The actual metadata is encoded in a 2:2 setup, that is, 2 data shards and 2 parity shards. This allows up to 2 (i.e. half) of the metadata stores to be lost, while still retaining access to the data. Any 2 stores can be used to recover the data, there is no specific difference between them.

Because the system is designed to prioritize recoverability over availability, writers will be rejected if the metadata storage is in the degraded state, that is, not all 4 stores are available and writeable.

A metadata store can be replaced by a new one, by removing the old one in the config and inserting the new one. The repair subsystem will take care of rebulding the data, regenerating the shards, and storing the new shards on the new metatada store.