/pyzdb

A lightweight database with Python syntax queries, using ZeroMQ

Primary LanguagePython

pyzdb - a lightweight database with Python syntax queries, using ZeroMQ

Please note this project's name change from pydb to pyzdb.

pyzdb ("pies db") is a database for storing nested list and dict and allows Python syntax queries instead of some variation of SQL. A deliberate choice is made to make no optimization on the queries so you know exactly what paths queries take.

Installation

pyzdb depends on

Install with

pip install -r requirements.txt

Note that undoable is not yet on PyPI and is installed using the -e flag. Alternatively, it can be downloaded manually and put in the same directory as server.py and client.py.

Running

In one terminal, run

python server.py

In a different terminal, run

python client.py

to get a prompt to access the database

> db

Sample session

> db
{}
> db['x'] = 3
None
> db['x']
3
> db['l'] = range(10)
None
> db
{u'x': 3, u'l': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]}
> db['l'][4] = ['a', 'b', 'c']
None
> db.undo()
None
> db
{u'x': 3, u'l': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]}
> db.undo()
None
> db.redo()
None
> db.redo()
None
> db
{u'x': 3, u'l': [0, 1, 2, 3, [u'a', u'b', u'c'], 5, 6, 7, 8, 9]}
> [v for v in db['l'][4]]
[u'a', u'b', u'c']
> [k for k in db]
[u'x', u'l']
> db['l'][4].append('d')
> db.save()

The server can be stopped and restarted to continue the session (without needing to restart the client). The most common use in an application is to import and use the client while the server is still run the same way.

from pyzdb.client import client
db, socket = client()
db['x'] = 3
x = db['x']._run()
db['l'] = range(10)
odd_squares = [v*v for v in db['l'] if v%2]
db.save()

lock example

from pyzdb.client import client
db, socket = client()
db.lock()
x = db['x']._run()
db['x'] = x + 1
db.unlock()

(or simply db['x'] = db['x']._run() + 1 but this will also run two queries instead of one.)

Multiple read-only servers example

In one terminal

python server.py

In a second terminal

python server.py -ro

In a third terminal

python client.py tcp://localhost:5559 tcp://localhost:5561

And proceed as previous examples.

tcp://localhost:5559 is the read-write URI and tcp://localhost:5561 is the read-only URI.

Note that all reads gets data from some consistent version of the database that's not necessarily the latest version so something like this from the client is possible.

> db['x'] = 3
None
> db.save()
None
> db['x'] = 4
None
> db.save()
None
> db['x']
3

If no parameters are passed to client.py or only a single parameter (the read-write URI), the client will only connect to the read-write server. This is the intended method for using only one database server.

Intended use

One intended use is a single instance of the database server with any number of clients (for example a web server) running on the same machine. The total database size isn't too large and large chunks are stored externally in files and represented by File objects in the database.

Another possibility is to have a single read-write server with multiple read-only server. The read-only server reload the database from the filesystem so to update them, update the database file.

Architecture

Architecture

server.py (with no arguments) starts

  • one queued router-dealer for read-only servers
  • one queued router-dealer for a read-write server
  • a read-write reply (zmq.REP) server.

The router-dealers serializes (and queues) incoming requests and sends one request at a time to the reply server of the right type. The reply server handles the request and answers the client (of that request).

All messages are encoded in JSON.

This post discusses choices when multiple read-only servers were added.

Debugging

exec_client.py is provided to help debugging. All requests sent from the client are executed (exec in globals()) on the server. This feature should probably be disabled for any public-facing program (safest would be to delete it from server.py).

Features and non-features

Data type

All data is JSON encoded and decoded so only JSON-encodable data can be stored, although its possible to write your own encoder to support more types of data.

Connection type

In theory, this database could allow any type of connection that ZeroMQ allows but most tests were done using TCP.

Query syntax

As seen the above examples, the first argument returned by client is treated as the root of nested list and dict and regular Python syntax is used to describe the entries we want to read or modify from that root (such as db['l'][4]).

To read an entry that is not an iterator, an extra ._run() function needs to be called (such as db['l'][4]._run()). This is because of the lazy evaluation implemented so db['l'][4] never actually sends any requests.

Serializability

lock and unlock in allows a client to get exclusive access to the database. No writes or reads are possible from other clients in the meantime. The client program has to be written so the order in which other accesses are treated are unimportant.

No deadlock (dead client) checks are implemented.

No support for locking only part of the database is implemented although this could be implemented using lock and unlock.

db.lock()
if not db['locks'].get(('l', 4), False) and not db['locks'].get(('l'), False):
    db['locks'][('l', 4)] = True
else:
    db.unlock()
    # return and do something else meanwhile
db.unlock()
# Do stuff on db['l'][4]
db.lock()
assert(db['locks'].get(('l', 4), False))
db['locks'][('l', 4)] = False
db.unlock()

Disk storage

The database itself is stored as a single pickled file but large blobs can be stored on the filesystem and a path in the database. If the client is on the same

Large files are "transfered" using a File("/path/to/file") object. The file needs to already be on the server's filesystem before a File object is stored.

Rollback

Server.undo and Server.redo are available but need to be called manually from a client.

Multiple database servers

Its now possible to have multiple read-only servers and one read-write server.

The read-only servers should reload the database from disk periodically (with the database file made available to them through some other means like a network filesystem or simply copying). No system for sending just the changes are available out of the box.

Authentication

There is no authentication mechanism. The intended usage is to have appropriate firewall rules outside the database.

Discussion

Here is a discussion about adding multiple read-only servers

Other similar projects