nd-db

[com.luposlip/nd-db "0.5.0"]

BREAKING CHANGE!

This library has been renamed from ndjson-db to nd-db. This includes the primary namespace! Furthermore the library identifier now follows Clojars new reverse domain name requirement.

Newline Delimited (read-only) Databases!

Clojure library that treats lines in newline delimited (potentially humongous) files as simple (thus lightening fast) databases.

nd-db currently works with JSON documents in .ndjson files, and EDN documents in .ndedn.

Usage

A very tiny test database resides in resources/test/test.ndjson.

It contains the following 3 documents, that has "id" as their unique IDs:

{"id":1, "data": ["some", "semi-random", "data"]}
{"id":222, "data": 42}
{"id":333333,"data": {"datakey": "datavalue"}}

Create Database

Since version 0.2.0 you need to create a database var before you can use the database. Behind the scenes this creates an index for you in a background thread.

(def db (nd-db.core/db {:id-fn #(Integer. ^String (second (re-find #"^\{\"id\":(\d+)" %))))
                        :filename "resources/test/test.ndjson"}))

If you want a default :id-fn created for you, use the :id-name together with :id-type and/or :source-type. Both :id-type and :source-type can be :string or :integer. :id-type is the target type of the indexed ID, whereas :source-type is the type in the source .ndjson database file. :source-type defaults to :id-type, and :id-type defaults to :string:

(def db (nd-db.core/db {:id-name "id"
                        :id-type :integer
                        :filename "resources/test/test.ndjson"}))

EDN

If you want to read a database of EDN documents, just use :doc-type :edn. Please note that the standard :id-name and :id-type parameters doesn't (as of v0.3.0) work with EDN, hence you need to implement the :id-fn accordingly.

Query Single Document

With a reference to the db you can query the database.

To find the data for the document with ID 222, you can perform a query-single:

(nd-db.core/q db 222)

Query Multiple Documents

You can also perform multiple queries at once. This is ideal for a pipelined scenario, since the return value is a lazy seq:

(nd-db.core/q db [333333 1 77])

It keeps!

NB: The above query for multiple documents, returns only 2 documents, since there is no document with ID 77. This is a design decision, as the documents themselves still contain the ID.

In a pipeline you'll be able to give lots of IDs to q, and filter down on documents that are actually represented in the database.

If you want to have an option to return nil in this case, let me know by creating an issue (or a PR).

The ID function

The ID functions adds "unlimited" flexibility as how to uniquely identify each document. You can choose a single attribute, or a composite. It's entirely up to you when you implement your ID function.

In the example above, the value of "id" is used as a unique ID to built up the database index.

Parsing JSON documents

If you use very large databases, it makes sense to think about performance in your ID function. In the above example a regular expression is used to find the value of "id", since this is faster than parsing JSON objects to EDN and querying them as maps.

Return value

Furthermore the return value of the function is (almost) the only thing being stored in memory. Because of that you should opt for as simple data values as possible. In the above example this is the reason for the parsing to Integer instead of keeping the String value.

Also note that the return value is the same you should use to query the database. Which is why the input to q are Integer instances.

Refer to the test for more details.

Persisting the database index

From v0.5.0 the generated index will be persisted to the temporary system folder on disk. This is a huge benefit if you need to use the same database multiple times, after throwing away the reference to the parsed database, since it takes much less time to read in the index as compared to parsing the database file.

For small files (like the sample databases found in this repository) it doesn't really make a difference. But for huge files, it makes an immense difference. The bigger the databases, the bigger the individual documents and the more complex the parsing of these documents are (to find the unique ID), the bigger the difference. For a .database file of 4.7GB the difference is 47s vs 90ms, or ~500 times faster!!

If you want to keep the serialized index file (*.nddbmeta) between system reboots, you should move it to another folder. You do that by using the parameter :index-folder to the db function.

If for some reason you don't want to persist the index - e.g. there's no storage attached to a docker container or serverless system - you can inhibit the persistence by setting param :index-persist? to false.

For more information on these and other parameters, see the source code for the db function in the core namespace.

Real use case: Verified Twitter Accounts

To test with a real database, download all verified Twitter users from here: https://files.pushshift.io/twitter/TU_verified.ndjson.xz

Put the file somewhere, i.e. path/to/TU_verified.ndjson, and run the following in a repl:

(time 
   (def katy-gaga-gates-et-al
     (doall
      (nd-db.core/q
       (nd-db.core/db {:id-name "screen_name" 
                       :filename "path/to/TU_verified.ndjson"})
       ["katyperry" "ladygaga" "BillGates" "ByMikeWilson"]))))

Performance

The extracted .ndjson files is 513 MB (297,878 records).

On my laptop the initial build of the index takes around 3 seconds, and the subsequent query of the above 3 verified Twitter users takes around 1 millisecond (specs: Intel® Core™ i7-8750H CPU @ 2.20GHz × 6 cores with 31,2 GB RAM, SSD HD).

In real usage scenarios, I've used 2 databases simultaneously of sizes 1.6 GB and 43.0 GB, with no problem or performance penalties at all (except for the relatively small size of the in-memory indices of course). Indexing the biggest database of 43GB took less than 2 minutes (NB: This is with a single core, and BEFORE 0.4.0).

Since the database uses disk random access, SSD speed up the database significantly.

Copyright & License

Copyright (C) 2019-2021 Henrik Mohr

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.