SIT is a lightweight TCP server that provides real-time full-text search over streams of JSON documents. It's also usable as a C library, where it can parse any stream using custom parsers.
- Lightweight The executable is 228K on my machine. It only depends on libev and libc; and currently (and temporarily) uses Ruby for the build and test suites.
- True realtime This is "true" real time, as opposed to "near" real time. SIT eliminates garbage collection, "commits," "flushes," and fsyncs. This is search highly optimized for recency and real-time. (However, we have not gone as far as hardware realtime or kernel guarantees)
- Write, then read Trivially support search results which include the document you just added.
- Computationally efficient percolation SIT is designed to add documents and execute query callbacks efficiently with up to 100,000 registered queries and/or connections.
- Flexible pubsub Rather than publishing into channels, you can use SIT to publish into a global event stream, and subscribe to dynamic channels defined by search queries.
SIT speaks a simple line-based pipelineable protocol. Any line that starts with
a curly brace ({
) is interpreted as a JSON document to add to the indexed
dataset. Other lines are interpreted as commands.
All responses are JSON lines, and have a status
key, which can be either ok,
or error. Other keys are implemented on a per-command basis.
SIT provides traditional search, where you give a query and get a resultset back. It also provides percolation, where you can register a query that will notify you when matching documents are added to the index.
-
register QUERY
Registers the query, or queries, for percolation. When any document that matches the query is added, SIT will print a "found" response in this stream.Sample request/response:
> register title ~ "hello world" AND points > 4; < {"status": "ok", "message": "registered", "id": 29} # ... < {"status": "ok", "message": "found", "query_id": 29, "doc_id": 500, "doc": {"title": "hello sweet world", "points": 7}}
The response to a
register
command includes an ID of the registered query. -
unregister QUERYID
Give the id provided in the register response, to stop the percolation.Sample request/response:
> unregister 29 < {"status": "ok", "message": "unregistered", "id": 29}
-
query QUERY
Do a search.Sample request/response:
# adding some docs > {"hello":"world 0"} > {"hello":"world 1"} > {"hello":"world 2"} > {"hello":"world 3"} > {"hello":"world 4"} > {"hello":"world 5"} > {"hello":"world 6"} > {"hello":"world 7"} > {"hello":"world 8"} > {"hello":"world 9"} > > query hello ~ world LIMIT 5; < {"status": "ok", "message": "querying", "id": 27} < {"status": "ok", "message": "found", "query_id": 27, "doc_id": 9, "doc": {"hello":"world 9"}} < {"status": "ok", "message": "found", "query_id": 27, "doc_id": 8, "doc": {"hello":"world 8"}} < {"status": "ok", "message": "found", "query_id": 27, "doc_id": 7, "doc": {"hello":"world 7"}} < {"status": "ok", "message": "found", "query_id": 27, "doc_id": 6, "doc": {"hello":"world 6"}} < {"status": "ok", "message": "found", "query_id": 27, "doc_id": 5, "doc": {"hello":"world 5"}} < {"status": "ok", "message": "complete", "id": 27}
SIT has a simple query language, composed of boolean operations (AND
, OR
,
NOT
) over clauses. You can append LIMIT N
to the end of a query. Queries
are terminated with either a newline or a semicolon. The following are valid
clauses:
field_name ~ string
field_name > integer
field_name = integer
field_name < integer
field_name >= integer
field_name <= integer
field_name != integer
SIT is designed to use pluggable tokenization strategies.
TODO: Describe tokenization 101 basics, concepts, examples.
- whitespace
- Unicode UAX#29 word boundaries
- regular expression pattern tokenizer
Pull requests are welcome.
The tilde indicates a full-text search. A full-text search identifies documents where a given term is present in the specified field.
The full-text search, title ~ hello
will match JSON documents with a field
named title
which contain a token of hello
. For example, the document
{"title":"hello full text search"}
when tokenized with a simple whitespace
tokenizer.
A quoted value for the tilde operator will match documents where all the terms
of the quoted text are present in the named field. For example, title ~ "hello world"
will be transformed into (title ~ hello AND title ~ world)
.
- Nested JSON support
- Composable streams
- Persistance
- Replace
- wchar/unicode support
- Facets
- Aggregations
- Quoted/span/proximity queries
- selectable parsers besides JSON in the server? (e.g., log output, msgpack, BSON)
Feeling adventurous? Run SIT on your own system and try some of these demos.
git clone git:https://github.com/fizx/sit.git
cd sit
bundle install
bundle exec rake
Start the server with ./sit
, which reads commands from standard input, and
prints its output to standard out.
./sit
{"hello":"world"}
# {"status": "ok", "message": "added", "doc_id": 0"}
query hello ~ world;
# {"status": "ok", "message": "querying", "id": 0}
# {"status": "ok", "message": "found", "query_id": 0, "doc_id": 0, "doc":
# {"hello":"world"}}
# {"status": "ok", "message": "complete", "id": 0}
Reminder, this is pre-release software. We use Twitter Streaming APIs as testing to find new and novel edge cases which cause crashes.
You should assume that this demo has a 50/50 chance to format your hard drive, and proceed accordingly (you're piping raw internet data through a pre-release C program).
First, install and authenticate twurl.
Next, start a server listening to the network.
./sit --port 4000
[INFO] [2013:03:0221:15:53] Successfully started server.
Now stream documents from Twitter, via netcat, to your running server.
twurl -t -H stream.twitter.com /1/statuses/sample.json | nc localhost 4000