site/src/Technique/Développement/Garage.md

# Garage: a S3-like object storage

Store pile of bytes in your garage.


## Context

Data storage is critical: it can lead to data loss if done badly and/or on hardware failure.
Filesystems + RAID can help on a single machine but a machine failure can put the whole storage offline.
Moreover, it put a hard limit on scalability. Often this limit can be pushed back far away by buying expensive machines.
But here we consider non specialized off the shelf machines that can be as low powered and subject to failures as a raspberry pi.

Distributed storage may help to solve both availability and scalability problems on these machines.
Many solutions were proposed, they can be categorized as block storage, file storage and object storage depending on the abstraction they provide. 

## Related work

Block storage is the most low level one, it's like exposing your raw hard drive over the network. 
It requires very low latencies and stable network, that are often dedicated. 
However it provides disk devices that can be manipulated by the operating system with the less constraints: it can be partitioned with any filesystem, meaning that it supports even the most exotic features.  
We can cite [iSCSI](https://en.wikipedia.org/wiki/ISCSI) or [Fibre Channel](https://en.wikipedia.org/wiki/Fibre_Channel).
Openstack Cinder proxy previous solution to provide an uniform API.

File storage provides a higher abstraction, they are one filesystem among others, which means they don't necessarily have all the exotic features of every filesystem.
Often, they relax some POSIX constraints while many applications will still be compatible without any modification.
As an example, we are able to run MariaDB (very slowly) over GlusterFS...
We can also mention CephFS (read [RADOS](https://ceph.com/wp-content/uploads/2016/08/weil-rados-pdsw07.pdf) whitepaper), Lustre, LizardFS, MooseFS, etc.
OpenStack Manila proxy previous solutions to provide an uniform API.

Finally object storages provide the highest level abstraction.
They are the testimony that the POSIX filesystem API is not adapted to distributed filesystems. 
Especially, the strong concistency has been dropped in favor of eventual consistency which is way more convenient and powerful in presence of high latencies and unreliability.
We often read about S3 that pioneered the concept that it's a filesystem for the WAN.
Applications must be adapted to work for the desired object storage service.
Today, the S3 HTTP REST API acts as a standard in the industry.
However, Amazon S3 source code is not open but alternatives were proposed.
We identified Minio, Pithos, Swift and Ceph.
Minio/Ceph enforces a total order, so properties similar to a (relaxed) filesystem.
Swift and Pithos are probably the most similar to AWS S3 with their consistent hashing ring.

There was many attempts in research too. I am only thinking to [LBFS](https://pdos.csail.mit.edu/papers/lbfs:sosp01/lbfs.pdf) that was used as a basis for Seafile.

- Cassandra (ScyllaDB) for metadata
- Own system using consistent hashing for data chunks


**Quentin:**

- pas d'erasure coding mais des checksums à côté des fichiers (ou dans les meta données)
- 2 ou 3 copies, configurable, potentiellement on a per bucket or per file basis
- on ne setup pas à la main en effet, je pensais au système qui scan sa partition de stockage et qui fait stockage géré = min(stockage de la partition - stockage que je ne gère pas, stockage alloué)
- La DHT/Ring à la dynamo, on doit pouvoir repomper un millier de truc sur leur papier. Surtout que je me le suis déja cogné deux fois. Le nombre d'entrées que tu mets est un multiple de ton stockage. eg: 500 Go, on fait des tranches de 10 Go --> 50 entrées dans le ring.
- un protocole de maintenance pompé sur le papier dynamo avec de l'anti entropy qui vérifie les blobs et leurs checksums (en plus de la vérification réalisée à la lecture)
- une interface web qui te donne en presque direct la santé de ton cluster (noeuds en vie, états de la réplication des données, problèmes de checksums)

**Other ideas:**

- split objects in constant size blocks or use the SeaFile strategy for better de-duplication? (Content Defined Chunking, Rabin's algorithm etc)

_Remark 1_ I really like the Rabin fingerprinting approach however deduplication means we need to implement reference counting. How do we implement it? If we suppose a CRDT counter, if we do +1, +1, -1 but counter is registered as +1, -1, +1, we are at zero at one point and lost ou chunk. ---> we need to be careful in our implementation if we want to play.

_Remark 2_ Seafile idea has been stolen from this article: https://pdos.csail.mit.edu/papers/lbfs:sosp01/lbfs.pdf

#### Random notes

--> we should not talk about block. It is the abstraction that manipulate your FS to interact with your hard drive. "Chunk" is probably more appropriate.  Block storage are a class of distributed storage where you expose the abstraction of your hard drive over the network, mainly SATA over ethernet, thinking to SCSI, FiberChannel, and so on

### Questions à résoudre


 1. est-ce que cassandra support de mettre certaines tables sur un SSD et d'autres sur un disque rotatif ?
 2. est-ce que cassandra/scylladb a un format de table on disk qui ne s'écroule pas complètement losque tu as des gros blobs ? (les devs de sqlite ont écrit tout un article pour dire que même avec leur lib qui est quand même sacrément optimisés, ils considèrent qu'à partir de je crois 4ko c'est plus efficace de mettre les blobs dans des fichiers séparés) - https://www.sqlite.org/intern-v-extern-blob.html
 3. Quelle taille de blocs ? L'idée c'est qu'on a quand même des liens en WAN avec des débits pas forcéments incroyables. Et ça serait bien que le temps de répliquer un bloc soit de l'ordre de la seconde maxi. En cas de retry, pour pouvoir mieux monitorer la progression, etc. Exoscale utilise 16Mo. LX propose 1Mo. 





#### Modules

- `membership/`: configuration, membership management (gossip of node's presence and status), ring generation --> what about Serf (used by Consul/Nomad) : https://www.serf.io/? Seems a huge library with many features so maybe overkill/hard to integrate
- `metadata/`: metadata management
- `blocks/`: block management, writing, GC and rebalancing
- `internal/`: server to server communication (HTTP server and client that reuses connections, TLS if we want, etc)
- `api/`: S3 API
- `web/`: web management interface


#### Metadata tables

**Objects:**

- *Hash key:* Bucket name (string)
- *Sort key:* Object key (string)
- *Sort key:* Version timestamp (int)
- *Sort key:* Version UUID (string)
- Complete: bool
- Inline: bool, true for objects < threshold (say 1024)
- Object size (int)
- Mime type (string)
- Data for inlined objects (blob)
- Hash of first block otherwise (string)

*Having only a hash key on the bucket name will lead to storing all file entries of this table for a specific bucket on a single node. At the same time, it is the only way I see to rapidly being able to list all bucket entries...*

**Blocks:**

- *Hash key:* Version UUID (string)
- *Sort key:* Offset of block in total file (int)
- Hash of data block (string)

A version is defined by the existence of at least one entry in the blocks table for a certain version UUID.
We must keep the following invariant: if a version exists in the blocks table, it has to be referenced in the objects table.
We explicitly manage concurrent versions of an object: the version timestamp and version UUID columns are index columns, thus we may have several concurrent versions of an object.
Important: before deleting an older version from the objects table, we must make sure that we did a successfull delete of the blocks of that version from the blocks table.

Thus, the workflow for reading an object is as follows:

1. Check permissions (LDAP)
2. Read entry in object table. If data is inline, we have its data, stop here.
   -> if several versions, take newest one and launch deletion of old ones in background
3. Read first block from cluster. If size <= 1 block, stop here.
4. Simultaneously with previous step, if size > 1 block: query the Blocks table for the IDs of the next blocks
5. Read subsequent blocks from cluster

Workflow for PUT:

1. Check write permission (LDAP)
2. Select a new version UUID
3. Write a preliminary entry for the new version in the objects table with complete = false
4. Send blocks to cluster and write entries in the blocks table
5. Update the version with complete = true and all of the accurate information (size, etc)
6. Return success to the user
7. Launch a background job to check and delete older versions


Workflow for DELETE:

1. Check write permission (LDAP)
2. Get current version (or versions) in object table
3. Do the deletion of those versions NOT IN A BACKGROUND JOB THIS TIME
4. Return succes to the user if we were able to delete blocks from the blocks table and entries from the object table

To delete a version:

1. List the blocks from Cassandra
2. For each block, delete it from cluster. Don't care if some deletions fail, we can do GC.
3. Delete all of the blocks from the blocks table
4. Finally, delete the version from the objects table


Known issue: if someone is reading from a version that we want to delete and the object is big, the read might be interrupted. I think it is ok to leave it like this, we just cut the connection if data disappears during a read.

("Soit P un problème, on s'en fout est une solution à ce problème")


#### Block storage on disk

**Blocks themselves:**

- file path = /blobs/(first 3 hex digits of hash)/(rest of hash)

**Reverse index for GC & other block-level metadata:**
 
- file path = /meta/(first 3 hex digits of hash)/(rest of hash)
- map block hash -> set of version UUIDs where it is referenced

Usefull metadata:

- list of versions that reference this block in the Casandra table, so that we can do GC by checking in Cassandra that the lines still exist
- list of other nodes that we know have acknowledged a write of this block, usefull in the rebalancing algorithm

Write strategy: have a single thread that does all write IO so that it is serialized (or have several threads that manage independent parts of the hash space). When writing a blob, write it to a temporary file, close, then rename so that a concurrent read gets a consistent result (either not found or found with whole content).

Read strategy: the only read operation is get(hash) that returns either the data or not found (can do a corruption check as well and return corrupted state if it is the case). Can be done concurrently with writes.

**Internal API:**

- get(block hash) -> ok+data/not found/corrupted
- put(block hash & data, version uuid + offset) -> ok/error
- put with no data(block hash, version uuid + offset) -> ok/not found plz send data/error
- delete(block hash, version uuid + offset) -> ok/error

GC: when last ref is deleted, delete block.
Long GC procedure: check in Cassandra that version UUIDs still exist and references this block.

Rebalancing: takes as argument the list of newly added nodes.

- List all blocks that we have. For each block:
- If it hits a newly introduced node, send it to them.
  Use put with no data first to check if it has to be sent to them already or not.
  Use a random listing order to avoid race conditions (they do no harm but we might have two nodes sending the same thing at the same time thus wasting time).
- If it doesn't hit us anymore, delete it and its reference list.

Only one balancing can be running at a same time. It can be restarted at the beginning with new parameters.


#### Membership management

Two sets of nodes:

- set of nodes from which a ping was recently received, with status: number of stored blocks, request counters, error counters, GC%, rebalancing%
  (eviction from this set after say 30 seconds without ping)
- set of nodes that are part of the system, explicitly modified by the operator using the web UI (persisted to disk),
  is a CRDT using a version number for the value of the whole set

Thus, three states for nodes:

- healthy: in both sets
- missing: not pingable but part of desired cluster
- unused/draining: currently present but not part of the desired cluster, empty = if contains nothing, draining = if still contains some blocks

Membership messages between nodes:

- ping with current state + hash of current membership info -> reply with same info
- send&get back membership info (the ids of nodes that are in the two sets): used when no local membership change in a long time and membership info hash discrepancy detected with first message (passive membership fixing with full CRDT gossip)
- inform of newly pingable node(s) -> no result, when receive new info repeat to all (reliable broadcast)
- inform of operator membership change -> no result, when receive new info repeat to all (reliable broadcast)

Ring: generated from the desired set of nodes, however when doing read/writes on the ring, skip nodes that are known to be not pingable.
The tokens are generated in a deterministic fashion from node IDs (hash of node id + token number from 1 to K).
Number K of tokens per node: decided by the operator & stored in the operator's list of nodes CRDT. Default value proposal: with node status information also broadcast disk total size and free space, and propose a default number of tokens equal to 80%Free space / 10Gb. (this is all user interface)


#### Constants

- Block size: around 1MB ? --> Exoscale use 16MB chunks
- Number of tokens in the hash ring: one every 10Gb of allocated storage
- Threshold for storing data directly in Cassandra objects table: 1kb bytes (maybe up to 4kb?)
- Ping timeout (time after which a node is registered as unresponsive/missing): 30 seconds
- Ping interval: 10 seconds
- ??


#### Links

 - CDC: <https://www.usenix.org/system/files/conference/atc16/atc16-paper-xia.pdf>
 - Erasure coding: <http://web.eecs.utk.edu/~jplank/plank/papers/CS-08-627.html>
 - [Openstack Storage Concepts](https://docs.openstack.org/arch-design/design-storage/design-storage-concepts.html)
 - [RADOS](https://ceph.com/wp-content/uploads/2016/08/weil-rados-pdsw07.pdf)