garage/doc/drafts/k2v-spec.md

Specification of the Garage K2V API (K2V = Key/Key/Value)

• We are storing triplets of the form (partition key, sort key, value) -> no user-defined fields, the client is responsible of writing whatever he wants in the value (typically an encrypted blob). Values are binary blobs, which are always represented as their base64 encoding in the JSON API. Partition keys and sort keys are utf8 strings.

• Triplets are stored in buckets; each bucket stores a separate set of triplets

• Bucket names and access keys are the same as for accessing the S3 API

• K2V triplets exist separately from S3 objects. K2V triples don't exist for the S3 API, and S3 objects don't exist for the K2V API.

• Values stored for triples have associated causality information, that enables Garage to detect concurrent writes. In case of concurrent writes, Garage keeps the concurrent values until a further write supersedes the concurrent values. This is the same method as Riak KV implements. The method used is based on DVVS (dotted version vector sets), described in the paper "Scalable and Accurate Causality Tracking for Eventually Consistent Data Stores", as well as here

Data format

Triple format

Triples in K2V are constituted of three fields:

• a partition key (pk), an utf8 string that defines in what partition the triple is stored; triples in different partitions cannot be listed together, they must be the object of different ReadItem or ReadBatch queries

• a sort key (sk), an utf8 string that defines the index of the triple inside its partition; triples are uniquely idendified by their partition key + sort key

• a value (v), an opaque binary blob associated to the partition key + sort key; they are transmitted as binary when possible but in most case in the JSON API they will be represented as strings using base64 encoding; a value can also be null to indicate a deleted triple (a null value is called a tombstone)

Causality information

K2V supports storing several concurrent values associated to a pk+sk, in the case where insertion or deletion operations are detected to be concurrent (i.e. there is not one that was aware of the other, they are not causally dependant one on the other). In practice, it even looks more like the opposite: to overwrite a previously existing value, the client must give a "causality token" that "proves" (not in a cryptographic sense) that it had seen a previous value. Otherwise, the value written will not overwrite an existing value, it will just create a new concurrent value.

The causality token is a binary/b64-encoded representation of a context, specified below.

A set of concurrent values looks like this:

(node1, tdiscard1, (v1, t1), (v2, t2)) ; tdiscard1 < t1 < t2
(node2, tdiscard2, (v3, t3)            ; tdiscard2 < t3

tdiscard for a node i means that all values inserted by node i with times <= tdiscard are obsoleted, i.e. have been read by a client that overwrote it afterwards.

The associated context would be the following: [(node1, t2), (node2, t3)], i.e. if a node reads this set of values and inserts a new values, we will now have tdiscard1 = t2 and tdiscard2 = t3, to indicate that values v1, v2 and v3 are obsoleted by the new write.

Basic insertion. To insert a new value v4 with context [(node1, t2), (node2, t3)], in a simple case where there was no insertion in-between reading the value mentionned above and writing v4, and supposing that node2 receives the InsertItem query:

• node2 generates a timestamp t4 such that t4 > t3.
• the new state is as follows:

(node1, tdiscard1', ())       ; tdiscard1' = t2
(node2, tdiscard2', (v4, t4)) ; tdiscard2' = t3

A more complex insertion example. In the general case, other intermediate values could have been written before v4 with context [(node1, t2), (node2, t3)] is sent to the system. For instance, here is a possible sequence of events:

1. First we have the set of values v1, v2 and v3 described above. A node reads it, it obtains values v1, v2 and v3 with context [(node1, t2), (node2, t3)].

2. A node writes a value v5 with context [(node1, t1)], i.e. v5 is only a successor of v1 but not of v2 or v3. Suppose node1 receives the write, it will generate a new timestamp t5 larger than all of the timestamps it knows of, i.e. t5 > t2. We will now have:

(node1, tdiscard1'', (v2, t2), (v5, t5)) ; tdiscard1'' = t1 < t2 < t5
(node2, tdiscard2, (v3, t3)              ; tdiscard2 < t3

3. Now v4 is written with context [(node1, t2), (node2, t3)], and node2 processes the query. It will generate t4 > t3 and the state will become:

(node1, tdiscard1', (v5, t5))       ; tdiscard1' = t2 < t5
(node2, tdiscard2', (v4, t4))       ; tdiscard2' = t3

Generic algorithm for handling insertions: A certain node n handles the InsertItem and is responsible for the correctness of this procedure.

1. Lock the key (or the whole table?) at this node to prevent concurrent updates of the value that would mess things up
2. Read current set of values
3. Generate a new timestamp that is larger than the largest timestamp for node n
4. Add the inserted value in the list of values of node n
5. Update the discard times to be the times set in the context, and accordingly discard overwritten values
6. Release lock
7. Propagate updated value to other nodes
8. Return to user when propagation achieved the write quorum (propagation to other nodes continues asynchronously)

Encoding of contexts:

Contexts consist in a list of (node id, timestamp) pairs. They are encoded in binary as follows:

checksum: u64, [ node: u64, timestamp: u64 ]*

The checksum is just the XOR of all of the node IDs and timestamps.

Once encoded in binary, contexts are written and transmitted in base64.

Indexing

K2V keeps an index, a secondary data structure that is updated asynchronously, that keeps tracks of the number of triples stored for each partition key. This allows easy listing of all of the partition keys for which triples exist in a bucket, as the partition key becomes the sort key in the index.

How indexing works:

• Each node keeps a local count of how many items it stores for each partition, in a local Sled tree that is updated atomically when an item is modified.
• These local counters are asynchronously stored in the index table which is a regular Garage table spread in the network. Counters are stored as LWW values, so basically the final table will have the following structure:

- pk: bucket
- sk: partition key for which we are counting
- v:  lwwmap (node id -> number of items)

The final number of items present in the partition can be estimated by taking the maximum of the values (i.e. the value for the node that announces having the most items for that partition). In most cases the values for different node IDs should all be the same; more precisely, three node IDs should map to the same non-zero value, and all other node IDs that are present are tombstones that map to zeroes. Note that we need to filter out values from nodes that are no longer part of the cluster layout, as when nodes are removed they won't necessarily have had the time to set their counters to zero.

API Endpoints

Operations on single items

ReadItem: GET /<bucket>/<partition key>?sort_key=<sort key>

Query parameters:

name	default value	meaning
sort_key	mandatory	The sort key of the item to read

Returns the item with specified partition key and sort key. Values can be returned in either of two ways:

1. a JSON array of base64-encoded values, or null's for tombstones, with header Content-Type: application/json

2. in the case where there are no concurrent values, the single present value can be returned directly as the response body (or an HTTP 204 NO CONTENT for a tombstone), with header Content-Type: application/octet-stream

The choice between return formats 1 and 2 is directed by the Accept HTTP header:

• if the Accept header is not present, format 1 is always used

• if Accept contains application/json but not application/octet-stream, format 1 is always used

• if Accept contains application/octet-stream but not application/json, format 2 is used when there is a single value, and an HTTP error 409 (HTTP 409 CONFLICT) is returned in the case of multiple concurrent values (including concurrent tombstones)

• if Accept contains both, format 2 is used when there is a single value, and format 1 is used as a fallback in case of concurrent values

• if Accept contains none, HTTP 406 NOT ACCEPTABLE is raised

Example query:

GET /my_bucket/mailboxes?sort_key=INBOX HTTP/1.1

Example response:

HTTP/1.1 200 OK
X-Garage-Causality-Token: opaquetoken123
Content-Type: application/json

[
  "b64cryptoblob123",
  "b64cryptoblob'123"
]

Example response in case the item is a tombstone:

HTTP/1.1 200 OK
X-Garage-Causality-Token: opaquetoken999
Content-Type: application/json

[
  null
]

Example query 2:

GET /my_bucket/mailboxes?sort_key=INBOX HTTP/1.1
Accept: application/octet-stream

Example response if multiple concurrent versions exist:

HTTP/1.1 409 CONFLICT
X-Garage-Causality-Token: opaquetoken123
Content-Type: application/octet-stream

Example response in case of single value:

HTTP/1.1 200 OK
X-Garage-Causality-Token: opaquetoken123
Content-Type: application/octet-stream

cryptoblob123

Example response in case of a single value that is a tombstone:

HTTP/1.1 204 NO CONTENT
X-Garage-Causality-Token: opaquetoken123
Content-Type: application/octet-stream

InsertItem: PUT /<bucket>/<partition key>?sort_key=<sort_key>

Inserts a single item. This request does not use JSON, the body is sent directly as a binary blob.

To supersede previous values, the HTTP header X-Garage-Causality-Token should be set to the causality token returned by a previous read on this key. This header can be ommitted for the first writes to the key.

Example query:

PUT /my_bucket/mailboxes?sort_key=INBOX HTTP/1.1
X-Garage-Causality-Token: opaquetoken123

myblobblahblahblah

Example response:

HTTP/1.1 200 OK

DeleteItem: DELETE /<bucket>/<partition key>?sort_key=<sort_key>

Deletes a single item. The HTTP header X-Garage-Causality-Token must be set to the causality token returned by a previous read on this key, to indicate which versions of the value should be deleted. The request will not process if X-Garage-Causality-Token is not set.

Example query:

DELETE /my_bucket/mailboxes?sort_key=INBOX HTTP/1.1
X-Garage-Causality-Token: opaquetoken123

Example response:

HTTP/1.1 204 NO CONTENT

Operations on index

ReadIndex: GET /<bucket>?start=<start>&end=<end>&limit=<limit>

Lists all partition keys in the bucket for which some triplets exist, and gives for each the number of triplets (or an approximation thereof, this value is asynchronously updated, and thus eventually consistent).

Query parameters:

name	default value	meaning
start	null	First partition key to list, in lexicographical order
end	null	Last partition key to list (excluded)
limit	null	Maximum number of partition keys to list

The response consists in a JSON object that repeats the parameters of the query and gives the result (see below).

The listing starts at partition key start, or if not specified at the smallest partition key that exists. It returns partition keys in increasing order and stops when either of the following conditions is met:

1. if end is specfied, the partition key end is reached or surpassed (if it is reached exactly, it is not included in the result)

2. if limit is specified, limit partition keys have been listed

3. no more partition keys are available to list

In case 2, and if there are more partition keys to list before condition 1 triggers, then in the result more is set to true and nextStart is set to the first partition key that couldn't be listed due to the limit. In the first case (if the listing stopped because of the end parameter), more is not set and the nextStart key is not specified.

Example query:

GET /my_bucket HTTP/1.1

Example response:

HTTP/1.1 200 OK

{ 
  start: null,
  end: null,
  limit: null,
  partition_keys: [
    [ "keys", 3043 ],
    [ "mailbox:INBOX", 42 ],
    [ "mailbox:Junk", 2991 ],
    [ "mailbox:Trash", 10 ],
    [ "mailboxes", 3 ],
  ],
  more: false,
  nextStart: null,
} 

Operations on batches of items

InsertBatch: POST /<bucket>

Simple insertion and deletion of triplets. The body is just a list of items to insert in the following format: [ "<partition key>", "<sort key>", "<causality token>"|null, "<value>"|null ].

The causality token should be the one returned in a previous read request (e.g. by ReadItem or ReadBatch), to indicate that this write takes into account the values that were returned from these reads, and supersedes them causally. If the triple is inserted for the first time, the causality token should be set to null.

The value is expected to be a base64-encoded binary blob. The value null can also be used to delete the triple while preserving causality information: this allows to know if a delete has happenned concurrently with an insert, in which case both are preserved and returned on reads (see below).

Partition keys and sort keys are utf8 strings which are stored sorted by lexicographical ordering of their binary representation.

Example query:

POST /my_bucket HTTP/1.1

[
  [ "mailbox:INBOX", "001892831", "opaquetoken321", "b64cryptoblob321updated" ],
  [ "mailbox:INBOX", "001892912", null, "b64cryptoblob444" ],
  [ "mailbox:INBOX", "001892932", "opaquetoken654", null ],
]

Example response:

HTTP/1.1 200 OK

ReadBatch: POST /<bucket>?search, or alternatively
ReadBatch: SEARCH /<bucket>

Batch read of triplets in a bucket.

The request body is a JSON list of searches, that each specify a range of items to get (to get single items, set single_item to true). A search is a JSON struct with the following fields:

name	default value	meaning
partition_key	mandatory	The partition key in which to search
start	null	The sort key of the first item to read
end	null	The sort key of the last item to read (excluded)
limit	null	The maximum number of items to return
single_item	false	Whether to return only the item with sort key start
conflicts_only	false	Whether to return only items that have several concurrent values
tombstones	false	Whether or not to return tombstone lines to indicate the presence of old deleted items

For each of the searches, triplets are listed and returned separately. The semantics of start, end and limit is the same as for ReadIndex. The additionnal parameter single_item allows to get a single item, whose sort key is the one given in start. Parameters conflicts_only and tombstones control additional filters on the items that are returned.

The result is a list of length the number of searches, that consists in for each search a JSON object specified similarly to the result of ReadIndex, but that lists triples within a partition key.

The format of returned tuples is as follows: [ "<sort key>", "<causality token>", "<value1>", ...], with the following fields:

• sort key: any unicode string used as a sort key

• causality token: an opaque token served by the server (generally base64-encoded) to be used in subsequent writes to this key

• value: binary blob, always base64-encoded

• if several concurrent values exist, they are appended at the end

• in case of concurrent update and deletion, a null is added to the list of concurrent values

• if the tombstones query parameter is set to true, tombstones are returned for items that have been deleted (this can be usefull for inserting after an item that has been deleted, so that the insert is not considered concurrent with the delete). Tombstones are returned as tuples in the same format with only null values

Example query:

POST /my_bucket?search HTTP/1.1

[
  {
    partition_key: "mailboxes",
  },
  {
    partition_key: "mailbox:INBOX",
    start: "001892831",
    limit: 3,
  },
  {
    partition_key: "keys",
    start: "0",
    single_item: true,
  },
]

Example associated response body:

HTTP/1.1 200 OK

[
  {
    partition_key: "mailboxes",
    start: null,
    end: null,
    limit: null,
    conflicts_only: false,
    tombstones: false,
    single_item: false,
    items: [
      [ "INBOX", "opaquetoken123", "b64cryptoblob123", "b64cryptoblob'123" ],
      [ "Trash", "opaquetoken456", "b64cryptoblob456" ],
      [ "Junk", "opaquetoken789", "b64cryptoblob789" ],
    ],
    more: false,
    nextStart: null,
  },
  {
    partition_key: "mailbox::INBOX",
    start: "001892831",
    end: null,
    limit: 3,
    conflicts_only: false,
    tombstones: false,
    single_item: false,
    items: [
      [ "001892831", "opaquetoken321", "b64cryptoblob321" ],
      [ "001892832", "opaquetoken654", "b64cryptoblob654" ],
      [ "001892874", "opaquetoken987", "b64cryptoblob987" ], 
    ],
    more: true,
    nextStart: "001892898",
  },
  {
    partition_key: "keys",
    start: "0",
    end: null,
    conflicts_only: false,
    tombstones: false,
    limit: null,
    single_item: true,
    items: [
      [ "0", "opaquetoken999", "b64binarystuff999" ],
    ],
    more: false,
    nextStart: null,
  },
]

DeleteBatch: POST /<bucket>?delete

Batch deletion of triplets. The request format is the same for POST /<bucket>?search to indicate items or range of items, except that here they are deleted instead of returned, but only the fields partition_key, start, end, and single_item are supported. Causality information is not given by the user: this request will internally list all triplets and write deletion markers that supersede all of the versions that have been read.

This request returns for each series of items to be deleted, the number of matching items that have been found and deleted.

Example query:

POST /my_bucket?delete HTTP/1.1

[
  {
    partition_key: "mailbox:OldMailbox",
  },
  {
    partition_key: "mailbox:INBOX",
    start: "0018928321",
    single_item: true,
  },
]

Example response:

HTTP/1.1 200 OK

[
  {
    partition_key: "mailbox:OldMailbox",
    start: null,
    end: null,
    single_item: false,
    deleted_items: 35,    
  },
  {
    partition_key: "mailbox:INBOX",
    start: "0018928321",
    end: null,
    single_item: true,
    deleted_items: 1,
  },
]

Internals: causality tokens

The method used is based on DVVS (dotted version vector sets). See:

• the paper "Scalable and Accurate Causality Tracking for Eventually Consistent Data Stores"
• https://github.com/ricardobcl/Dotted-Version-Vectors

For DVVS to work, write operations (at each node) must take a lock on the data table.