pgcopy is a small system for very fast bulk insertion of data into a PostgreSQL database table using binary copy.
To install:
$ pip install pgcopy
pgcopy requires pytz and the psycopg2 db adapter. nose is required to run the tests.
pgcopy provides facility for copying data from an iterable of tuple-like
objects using a CopyManager, which must be instantiated with a psycopg2
db connection, the table name, and an iterable containing the names of the
columns to be inserted in the order in which they will be provided.
pgcopy inspects the database to determine the datatypes of the columns.
For example:
from datetime import datetime
from pgcopy import CopyManager
import psycopg2
cols = ('id', 'timestamp', 'location', 'temperature')
now = datetime.now()
records = [
(0, now, 'Jerusalem', 72.2),
(1, now, 'New York', 75.6),
(2, now, 'Moscow', 54.3),
]
conn = psycopg2.connect(database='weather_db')
mgr = CopyManager(conn, 'measurements_table', cols)
mgr.copy(records)
# don't forget to commit!
conn.commit()
By default, a temporary file on disk is used. If there's enough memory, you can get a slight performance benefit with in-memory storage:
from io import BytesIO mgr.copy(records, BytesIO)
A db schema can be specified in the table name using dot notation:
mgr = CopyManager(conn, 'myschema.measurements', cols)
Currently the following PostgreSQL datatypes are supported:
- bool
- smallint
- integer
- bigint
- real
- double precision
- char
- varchar
- text
- bytea
- date
- timestamp
- timestamp with time zone
- numeric (data must be
decimal.Decimal) - json
- jsonb
- uuid
Unicode strings in the data to be inserted (all values of type str in
Python 3) should be encoded as bytes before passing them to copy.
Values intended to be NULL in the database should be encoded as None
rather than as empty strings.
Note
PostgreSQL numeric does not support Decimal('Inf') or
Decimal('-Inf'). pgcopy serializes these as NaN.
For a fast test run using current environment, use nose:
$ nosetests
For more thorough testing, Tox configuration will run tests on python versions 2.7 and 3.3 - 3.6:
$ tox
Additionally, test can be run with no local requirements other than the ubiquitous docker:
$ docker-compose up pgcopy
Below are simple benchmarks for 100000 records. This gives a general idea of the kind of speedup available with pgcopy:
$ nosetests -c tests/benchmark.cfg
ExecuteManyBenchmark: 7.75s
PGCopyBenchmark: 0.54s
----------------------------------------------------------------------
Ran 2 tests in 9.101s
When possible, faster insertion may be realized by inserting into an empty
table with no indices or constraints. In a case where the entire contents
of the table can be reinserted, the Replace context manager automates
the process. On entry, it creates a new table like the original, with a
temporary name. Default column values are included. It provides the
temporary name for populating the table within the context. On exit, it
recreates the constraints, indices, triggers, and views on the new table,
then replaces the old table with the new. It can be used so:
from pgcopy import CopyManager, Replace
with Replace(conn, 'mytable') as temp_name:
mgr = CopyManager(conn, temp_name, cols)
mgr.copy(records)
Replace renames new db objects like the old, where possible.
Names of foreign key and check constraints will be mangled.
As of v0.6 there is also pgcopy.util.RenameReplace, which instead of
dropping the original objects renames them using a transformation function.
Note that on PostgreSQL 9.1 and earlier, concurrent queries on the table will fail once the table is dropped.
cpgcopy, a Cython implementation, about twice as fast.