Skip to content

Latest commit

 

History

History
173 lines (128 loc) · 5.76 KB

README.md

File metadata and controls

173 lines (128 loc) · 5.76 KB

flumeview-query

A flumeview with map-filter-reduce queries

Motivation

This particular module was because I needed to query things in secure-scuttlebutt in a flexible way. A previous exploration of this general idea was mynosql Yes the joke is that it's sql, but for no sql. SQL is actually totally functional (just with weird names)

map, filter, reduce == select, where, group-by !

Except with none of those icky schemas that just get you down! (anyway, with ssb we really can't enforce schemas (because of privacy oriented decentralization))

Example

var db = Flume(log).use('query', FlumeQuery(null, {indexes:indexes}))

//write a batch of data to the log
db.append([{
  foo: true,
  bar: 5,
  nested: {baz: 'okay'}
},
{
  foo: false,
  bar: 6,
  nested: {baz: 'okay'}
},
{
  foo: false,
  bar: 7,
  nested: {baz: 'not-okay'}
},
], function (err) {
  //filter for all records which match the above query
  pull(
    db.query.read({query: [
      {$filter: {nested: {baz: 'okay'}}}
    ]}),
    pull.collect(function (err, ary) {
      console.log(ary)
      //out puts the first and second items inserted above.
    })
  )
})

new api

flumeview-query shouldn't hold the indexes, it should just know what to do with them. Instead of passing in the indexes to create, pass in an index which can be used. that index would need to expose what paths it indexes, of course.

query.add(index, createStream) //add an index

Indexes

The indexes argument is an array of indexes that flumeview-query will be able to look at to do a fast query.

Indexes is an object with the a short name of the index (this will be stored with every record, so say 3 chars, is recommended) and a value with the fields being indexed.

For example, ssb-query's indexes look like:

[
  {key: 'log', value: ['timestamp']},
  {key: 'clk', value: [['value', 'author'], ['value', 'sequence']] },
  {key: 'typ', value: [['value', 'content', 'type'], ['timestamp']] },
  {key: 'cha', value: [['value', 'content', 'channel'], ['timestamp']] },
  {key: 'aty', value: [['value', 'author'], ['value', 'content', 'type'], ['timestamp']]}
]

Indexes can be of a single field or multiple fields (which are called "compound indexes"). Each item in an index must be unique, that is why the most of the indexes end in timestamp, author:sequence is also unique in ssb, so that index doesn't need timestamp. The uniqueness is not enforced by flumeview-query, it is the responsibilty of the index designer.

Compound indexes optimize queries with multiple fields. For example a query like: "all posts by @bob" which is {value: {author: @bob, content: {type: 'post'}}} uses the clk index.

If a query matches all fields in the index, the query will return 1 item (or zero if there is no record with those values), otherwise results will be returned in the order of the index.

Properties in the index are used from left to right, a query for "messages from @bob received since yesterday" {author: @bob, timestamp: {$gt: yesterday}} cannot use author and timestamp fields in aty as it leaves a gap in the value.content.type field. This query with these indexes would use part of a compound index, (clk, because it matches first) read all the messages by @bob and filter out the records matching the other query parameter (timestamp: {$gt: yesterday}). This is called a "partial scan". A partial scan is clearly less efficient than matched index, but not as bad as a full scan (which reads the entire database!)

Queries with compound indexes will end up sorted by the last index matched. Therefor, put the fields you expect to be exact first!

Queries

This module uses map-filter-reduce queries, if the filter stage uses fields that are in a index, then select can choose the best index and perform many queries very quickly.

See map-filter-reduce for documentation of the syntax, for example queries, performed on top of secure-scuttlebutt

api : flumedb.use("query", FlumeViewQuery(version, opts))

version must be an number. When you update any options, change the version and the index will rebuild. opts is the options. in particular {indexes: [...]} is mandatory.

Here we use the name "query", you can use any name.

api : FlumeViewQuery(version, {indexes, filter, map}) => FlumeView: query

as required by every flumeview, version is a integer. change this when the indexes or other settings change and the view will be rebuilt. indexes is mandatory. Indexes are the paths supported. indexes to use. links is an optional function used for mapping a value into one or more values for the index. version must be an number. When you change indexes or links, bump the version and the index will rebuild.

query.read({query:MFR_query, limit, reverse, live, old})

Perform the query! limit, reverse, live, old are stardard as with other flume streams. unlinkedValues is an option that can be used to include the values not part of the index in the return value.

query.explain ({query:MFR_query, limit, reverse, live, old}) => obj

Figure out what indexes are best to use to perform a query, but do not actually run the query! If a query is slow or doesn't seem to be working right, this method can be used to understand what is really going on. If the return value is {scan: true} that means no indexes are being used. If an index is selected, that should mean it's more efficient, but it might still be filtering the output.

License

MIT