dull

Playing around with multilevel-http, node-hashring and express-swim to build a riak-like distributed leveldb, any message passing between cluster nodes is http-based.

$ npm install -g dull
$ mkdir -p data/node1 data/node2 data/node3
$ dull --port 3001 --path ./data/node1
$ dull --port 3002 --path ./data/node2 --join 127.0.0.1:3001
$ dull --port 3003 --path ./data/node3 --join 127.0.0.1:3001
$ curl -X PUT -d '{ "cap": { "n": 3 }  }' http://localhost:3001/bucket/people
$ curl http://localhost:3001/buckets/data
$ curl -X PUT -d '{ "name": "Andrea", "lastname": "Gariboldi", age: 33 }' -H 'Content-Type: application/json' http://localhost:3002/bucket/people/data/andrea
$ curl http://localhost:3002/bucket/people/data/andrea
$ curl -X DELETE http://localhost:3002/bucket/people/data/andrea
$ mkdir data/node4
$ dull --port 3004 --path ./data/node4
$ curl -X POST -d '127.0.0.1:3001' http://localhost:3004/gossip/join
$ curl -X DELETE http://localhost:3004/gossip/leave
$ curl http://localhost:3002/bucket/people/keys
$ curl -X PUT --data-binary @examples/v8.png -H 'Content-Type: image/png' http://127.0.0.1:3001/bucket/people/data/v8.png

CLI arguments

node options

• --host <ip address> (default: 127.0.0.1): listen address.
• --port <port> (default: 3000): listen port.
• --path <dir> (default: ./data): directory to store data files.
• --timeout <ms> (default: 1000ms): timeout for request between nodes.
• --join <active node> (default: (empty)): another node already in the cluster to join it.

cap controls

• --cap.n <n> (default: 3): number of replicas for a given key.
• --cap.r <r> (default: 2): number of replicas that should respond to a read request, for it to be considered successful.
• --cap.w <w> (default: 2): number of replicas that should respond to a write request, for it to be considered successful.

vector clocks options

• --vclock.small <n> (default: 10): number of entries to consider a vector clock small.
• --vclock.big <n> (default: 50): number of entries to consider a vector clock big.
• --vclock.young <n> (default: 20): max number of seconds to consider a vector clock entry young.
• --vclock.old <n> (default: 86400): min number of seconds to consider a vector clock entry old.

ring options (hashring)

• --ring.vnode_count <n> (default: 40): The amount of virtual nodes per server.
• --ring.max_cache_size <n> (default: 5000): We use a simple LRU cache inside the module to speed up frequent key lookups, you can customize the amount of keys that need to be cached.

gossip options (express-swim)

• --swim.verbose (default: false): make the gossip protocol very verbose.
• --swim.period_length (default: 3000): period length (in milliseconds).
• --swim.ping_timeout (default: 1000): timeout of a ping request (in milliseconds).
• --swim.failing_timeout (default: 9000): timeout of a suspected state before failing a node (in milliseconds).
• --swim.message_ttl (default: 10000): the time a message should be kept in the message queue (in milliseconds).
• --swim.pingreq_nodes (default: 2): number of random nodes to select for a ping-req.
• --swim.tune_gossip (default: 2): tune maximum message retransmission (keep it "small").
• --swim.gossip_messages (default: 10): max piggybacked messages per request.

HTTP API

A node in dull is the string host:port.

Cluster (express-swim)

join a node to a cluster

curl -X POST -d <active node> http://<joining node>/gossip/join

make a node leave the cluster

curl -X DELETE http://<leaving node>/gossip/leave

list active nodes in the cluster

curl http://<active node>/gossip/nodes

Buckets (server/buckets.js)

create or update bucket and options

curl -X PUT -d <bucket options> http://<active node>/bucket/<bucket name>

delete a bucket

curl -X DELETE http://<active node>/bucket/<bucket name>

list buckets

curl http://<active node>/buckets/keys
curl http://<active node>/buckets/data
curl http://<active node>/buckets/values

Data (server/data.js)

put a KV

curl -X PUT -d <value> http://<active node>/bucket/<bucket name>/data/<key>

this defaults to Content-Type: application/json. You can put any other value type like this:

curl -X PUT -d <value> -H 'Content-Type: <value type>' http://<active node>/bucket/<bucket name>/data/<key>

or

curl -X PUT --data-binary @<file name> -H 'Content-Type: <value type>' http://127.0.0.1:3001/bucket/<bucket name>/data/<key>

get a value

curl -v http://<active node>/bucket/<bucket name>/data/<key>

you will get an header x-dull-vclock like this:

x-dull-vclock: {"127.0.0.1:3002":2,"127.0.0.1:3001":1}

that you should pass back updating a value like this:

curl -X PUT -d <value> -H 'Content-Type: <value type>' -H 'x-dull-vclock: {"127.0.0.1:3002":2,"127.0.0.1:3001":1}' http://<active node>/bucket/<bucket name>/data/<key>

those are used in read-repair.

delete a KV

curl -X DELETE http://<active node>/bucket/<bucket name>/data/<key>

you will get an header x-dull-vclock like this:

x-dull-vclock: {"127.0.0.1:3002":2,"127.0.0.1:3001":1}

that you should pass back if you want to recreate the value like this:

curl -X PUT -d <value> -H 'Content-Type: <value type>' -H 'x-dull-vclock: {"127.0.0.1:3002":2,"127.0.0.1:3001":1}' http://<active node>/bucket/<bucket name>/data/<key>

Deleted objects are marked with a thumbstone record that may be in conflict with a put if you don't pass in a vector-clock. This thumbstone record is also used when generating keys: dull will return only non-deleted keys as one would expect, resolving those vector clocks. It will return also keys that has conflicts, so that the client may have the chance to decide it.

list all keys known to dull

curl http://<active node>/bucket/<bucket name>/keys

increment/decrement a pn-counter

curl -X PUT -d <n> http://<active node>/counter/<bucket name>/data/<key>

get the counter value

curl http://<active node>/counter/<bucket name>/data/<key>

custom http headers

x-dull-clientid

Used to identify the actor in vector clocks, if no actor is passed in an uuid is used.

x-dull-vclock

used to pass vclocks between client and the server: the server will return a vector clock in read or delete of a key, that should be used to modify the key.

x-dull-thumbstone

it is used to identify a deleted key sibling on a 303 response.

Access to any node leveldb

You can check multilevel-http

Buckets

curl http://<active node>/buckets/<multilevel-http path>

Data

curl http://<active node>/mnt/<bucket name>/<multilevel-http path>

Things to implement

Vclock

http://basho.com/why-vector-clocks-are-hard/

• vclock header compression

CRDTs

http://vimeo.com/43903960 https://github.com/aphyr/meangirls http://docs.basho.com/riak/latest/dev/using/data-types/

Tests

https://github.com/aphyr/jepsen

Active anti-entropy

http://en.wikipedia.org/wiki/Merkle_tree https://github.com/c-geek/merkle

Hinted hand-off

leveldb range of keys to push to a node when it comes back