flumedb

A modular database made for moving logs with streams.

architecture

Flume is a modular database comprised of an Append Only Log and Streaming Views on that log. This makes a star shaped pipeline - or rather, there is a pipeline from the log to each view, but the log is part of every pipeline.

The Log is the main storage element and provides durability. Views stream that data and build up their own model. They can point back to the main storage (normalized) or materialize it (denormalized). The views can use a structure optimized for the queries they provide, instead of durability, because the can rebuild from the main log.

In my previous modular database design, level-sublevel indexes had to be written atomically with the data. This was okay for map style indexes, but made simple aggregations such as a count of all records difficult. Worse, rebuilding the indexes needed a batch migration between two versions of the database. This made developing interesting applications on level-sublevel quite difficult (such as secure-scuttlebutt)

In flume, each view remembers a version number, and if the version number changes, it just rebuilds the view. This means view code can be easily updated, or new views added. It just rebuilds the view on startup. (though, this may take a few minutes on larger data)

The trick is that each view exposes a observable, flumeview.since, this represents it's current state - the sequence number the view is up to. An observable is like an event meets a value. it's an changing value that you can observe. Events, promises, or streams are similar, but not quite the right choice here.

Note, views are async. The main log may callback before the view is fully up to date, but if a read is made to an as yet unsynced view, it just waits for the view building to complete. This may make the call take longer (applications should show a progress bar for reindexing) but most of the application code can just assume that indexes are always up to date, because if you get a callback from an append to the log, a subsequent view read will return data consistent with that.

This gives us the freedom to have async views, which gives us the freedom to have many different sorts of views!

example

Take one flumelog-* module and zero or more flumeview-* modules, and glue them together with flumedb.

var MemLog = require('flumelog-memory') // just store the log in memory
var Reduce = require('flumeview-reduce') // just a reduce function.
var Flume = require('flumedb')

var db = Flume(MemLog())
  // the api of flumeview-reduce will be mounted at db.sum...
  .use(
    'sum',
    Reduce(1, function (acc, item) {
      return (acc || 0) + item.foo
    })
  )

db.append({ foo: 1 }, function (err, seq) {
  if (err) throw err
  db.sum.get(function (err, value) {
    if (err) throw err
    console.log(value) // 1
  })
})

modules

There are two types of components, logs, which you need just one of, and views, which you want many of. I expect that the list of view modules will grow much longer than the list of log modules.

logs

• flumedb/flumelog-offset a log in a file recommended.
• flumedb/flumelog-memory a log that is just an in memory array.
• flumedb/flumelog-level a log on level.
• flumedb/flumelog-idb flumelog on top of indexeddb.

views

• flumedb/flumeview-reduce a reduce function as a view.
• flumedb/flumeview-level an implemented index on level.
• flumedb/flumeview-query functional query language.
• flumedb/flumeview-search full text search.
• flumedb/flumeview-hashtable hashtable is ideal when you have uniqueish keys and do not need range queries.
• flumedb/flumeview-bloom bloom filter lets you check if may have something.

other

other modules that may come in handy

• flumedb/flumecodec simple codecs, used in most views and logs.
• dominictarr/charwise an order-preserving encoding compatible with bytewise, but faster.
• flumedb/flumecli instance command line interface
• flumedb/aligned-block-file read and write to a file with aligned blocks that are easy to cache.
• flumedb/test-flumelog reusable tests for a flumelog.

api

flumedb api

Flume(flumelog, isReady?, mapper?) => flumedb

construct a flumedb instance from a flumelog. if isReady is false, nothing will happen with the views until flumedb.ready.set(true) is called. This is useful if some sort of migration needs to happen before everything starts processing.

mapper is an optional async function function (value, cb) { cb(null, value) } that map apply some transformation onto a data item before passing it to the views, get or stream.

flumedb.get (seq, cb)

This exposes get from the underlying flumelog module. This takes a seq as is the keys in the log, and returns the value at that sequence or an error.

flumedb.stream (opts)

This exposes stream from the underlying flumelog module. It supports options familiar from leveldb ranges, i.e. lt, gt, lte, gte, reverse, live, limit.

flumedb.since => observable

The observable which represents the state of the log. This will be set to -1 if the view is empty, will monotonically increase as the log is appended to. The particular format of the number depends on the implementation of the log, but should usually be a number, for example, a incrementing integer, a timestamp, or byte offset.

flumedb.append (value, cb(err, seq))

Appends a value to the database. The encoding of this value will be handled by the flumelog, will callback when value is successfully written to the log (or it errors). On success, the callback will return the new maximum sequence in the log.

If value is an array of values, it will be treated as a batched write. By the time of the callback, flumedb.since will have been updated.

flumedb.use(name, createFlumeview) => self

Installs a flumeview module. This will add all methods that the flumeview describes in it's modules property to flumedb[name] If name is already a property of flumedb then an error will be thrown. You probably want to add plugins at startup, but given the streaming async nature of flumedb, it will work completely fine if you add them later!

flumedb.rebuild (cb)

Destroys all views and rebuilds them from scratch. If you have a large database and lots of views this might take a while, because it must read completely through the whole database again. cb will be called once the views are back up to date with the log. You can continue using the various read APIs and appending to the database while it is being rebuilt, but view reads will be delayed until those views are up to date.

flumedb.close(cb)

closes all flumeviews.

flumedb.closed => boolean

set to true if flumedb.close() has been called

flumedb.views => object

flumeviews that are currently being used by flumedb

flumelog api

The flume log is the heart of the flumedb setup - it's the cannonical store of data, and all state is stored there. In contrast, all flume views are derived / generated completely from the data in the log, so they can be blown away and regenerated easily.

Each type of log defines the following methods, which are then exposed by flumedb (e.g. flumelog.get is what defines flumedb.get) for you to access.

flumelog.get (seq, cb)

Retrives the value at position seq in the log.

flumelog.stream(opts)

Returns a source stream over the log. It supports options familiar from leveldb ranges, i.e. lt, gt, lte, gte, reverse, live, limit.

flumelog.since => observable

An observable which represents the state of the log. If the log is uninitialized it should be set to undefined, and if the log is empty it should be -1 if the log has data, it should be a number larger or equal to zero.

flumelog.append (value, cb(err, seq))

Appends a value (or array of values) to the log, and returns the new latest sequence. flumelog.since is updated before calling cb.

flumelog.dir => directory name

The filename of the directory this flumelog is persisted in (this is used by the views, if they are also persistent).

flumeview api

A flumeview provides a read API, and a streaming sink that accepts data from the log (this will be managed by flumedb).

flumeview.since => observable

The current state of the view (this must be the sequence of the last value processed by the view) a flumeview must process items from the main log in order, otherwise inconsistencies will occur.

flumeview.createSink (cb)

Returns a pull-stream sink that accepts data from the log. The input will be {seq, value} pairs. cb will be called when the stream ends (or errors).

flumeview.destroy (cb)

Wipe out the flumeview's internal (and persisted) state, and reset flumeview.since to -1.

flumeview.close (cb)

flush any pending writes and release resources, etc.

flumeview.methods = {<key>:'sync'|'async'|'source'}

An object describing the methods exposed by this view. A view needs to expose at least one method (otherwise, why is it useful?).

These the corresponding methods will be added at flumedb[name][key]. If they type is async or source the actual call to the flumeview[key] method will be delayed until flumeview.since is in up to date with the log. sync type methods will be called immediately.

If you would rather not wait for the view to get up to date with the log, but just want the data that is already in the index. provide the option since: -1 to the view call. This works for async and source apis. If since is a positive number, the view will wait until it's at least up that up to date.

flumeview.ready (cb)

A ready method is also added to each mounted flumeview which takes a callback which will be called exactly once, when that view is up to date with the log (from the point where it is called).

License

MIT