|Michael Ekstrand 5b49d61a89 Merge remote-tracking branch 'public/master'||3 weeks ago|
|.dvc||5 months ago|
|bookdata||3 weeks ago|
|data||3 weeks ago|
|docs||3 weeks ago|
|import||3 weeks ago|
|index||3 weeks ago|
|integrate||3 weeks ago|
|schemas||3 weeks ago|
|scripts||3 weeks ago|
|src||3 weeks ago|
|.editorconfig||4 months ago|
|.gitattributes||4 months ago|
|.gitignore||4 months ago|
|Cargo.lock||3 weeks ago|
|Cargo.toml||3 weeks ago|
|ClusterStats.ipynb||4 months ago|
|Dvcfile||3 weeks ago|
|IDGraphExplore.ipynb||4 months ago|
|LICENSE.txt||1 year ago|
|LinkageStats.ipynb||4 months ago|
|README.md||4 months ago|
|dvc.cmd||4 months ago|
|dvc.sh||5 months ago|
|environment.yml||3 weeks ago|
|init.dvc||3 weeks ago|
|ol-explore.sql||1 year ago|
|run.py||4 months ago|
|schemas.dvc||3 weeks ago|
|DVC Managed File|
|Git Managed File|
|DVC Managed File|
|Git Managed File|
This repository contains the code to import and integrate the book and rating data that we work with. It imports and integrates data from several sources in a single PostgreSQL database; import scripts are primarily in Python, with Rust code for high-throughput processing of raw data files.
If you use these scripts in any published research, cite our paper:
Michael D. Ekstrand, Mucun Tian, Mohammed R. Imran Kazi, Hoda Mehrpouyan, and Daniel Kluver. 2018. Exploring Author Gender in Book Rating and Recommendation. In Proceedings of the 12th ACM Conference on Recommender Systems (RecSys '18). ACM, pp. 242–250. DOI:10.1145/3240323.3240373. arXiv:1808.07586v1 [cs.IR].
Note: the limitations section of the paper contains important information about the limitations of the data these scripts compile. Do not use the gender information in this data data or tools without understanding those limitations. In particular, VIAF's gender information is incomplete and, in a number of cases, incorrect.
In addition, several of the data sets integrated by this project come from other sources with their own publications. If you use any of the rating or interaction data, cite the appropriate original source paper. For each data set below, we have provided a link to the page that describes the data and its appropriate citation.
pg_prewarm(from the PostgreSQL Contrib package) installed.
It is best if you do not store the data files on the same disk as your PostgreSQL database.
environment.yml file defines an Anaconda environment that contains all the required packages except for the PostgreSQL server. It can be set up with:
conda create -f environment.yml
We use Data Version Control (
dvc) to script the import and wire
its various parts together. A complete re-run, not including file download time, takes
approximately 8 hours on our hardware (24-core 2GHz Xeon, 128GiB RAM, spinning disks).
All scripts read database configuration from the
DB_URL environment variable, or alternately
a config file
db.cfg. This file should look like:
[DEFAULT] host = localhost database = bookdata
This file additionally supports branch-specfic configuration sections that will apply to work on different Git branches, e.g.:
[DEFAULT] host = localhost database = bookdata [master] database = bdorig
This setup will use
bookdata for most branches, but will connect to
bdorig when working
master branch in the git repository.
This file should not be committed to Git. It is ignored in
After creating your database, initialize the extensions (as the database superuser):
CREATE EXTENSION orafce; CREATE EXTENSION pg_prewarm; CREATE EXTENSION "uuid-ossp";
The default PostgreSQL performance configuration settings will probably not be very effective; we recommend turning on parallelism and increasing work memory, at a minimum.
This imports the following data sets:
data(not auto-downloaded - save CSV file in
data). If you use this data, cite the paper on that site.
data). If you use this data, cite the paper on that site.
Several of these files can be auto-downloaded with the DVC scripts; others will need to be manually downloaded.
You can run the entire import process with:
Individual steps can be run with their corresponding
The import code consists of Python, Rust, and SQL code, wired together with DVC.
Python scripts live under
scripts, as a Python package. They should not be launched directly, but
run.py, which will make sure the environment is set up properly for them:
python run.py sql-script [options] script.sql
Our SQL scripts are run with a custom SQL script runner (the
sql-script Python script), that breaks
them into chunks, handles errors, and tracks dependencies and script status. The script runner parses
directives in SQL comments; for example:
--- #step ISBN ID storage CREATE TABLE IF NOT EXISTS isbn_id ( isbn_id SERIAL PRIMARY KEY, isbn VARCHAR NOT NULL UNIQUE );
is a step called "ISBN ID storage". Each step is processed in a transaction that is committed at the
end, so steps are atomic (unless marked with
These are the directives for steps:
#step LABELstarts a new step with the label
LABEL. Additional directives before the first SQL statement will apply to this step.
#notxmeans the step will run in autocommit mode. This is needed for certain maintenance commands that do not work within transactions.
#allow CODEallows the PostgreSQL error 'code', such as
invalid_table_definition. The script will not fail if the step fails with this error. Used for dealing with steps that do things like create indexes, so if the index already exists it is fine to still run the script.
In addition, the top of the file can have
#dep directives, that indicate the dependencies of this
script. The only purpose of the
#dep is to record dependencies in the database stage state
table, so that modifications can propagate and be detected; dependencies still need to be recorded
.dvc files to run the import steps in the correct order.
Running the scripts here with raw
dvc does not work. You need to use the
script, as in:
The wrapper script sets up DVC to recognize our special
pgstat://stage URLs for tracking the
status of database import stages in the live database.
Import is structured as a concept of stages map almost 1:1 to our DVC step files. They manage database-side tracking of data and status.
Each import stage includes
pgstat://stage as an unached output stage, as in:
outs: - path: pgstat://bx-import cache: false
From the command line, uncached outptus are created by using
-O instead of
Each script that requires another stage to be run first depends on
pgstat://stage as a dependency.
This wires together all of the dependencies, and uses the current state in the database instead of files that might become out-of-sync with the database to track import status.
The stage name matches the name of the
The reason for this somewhat bizarre layoutis that if we just wrote the output files, and the database was reloaded or corrupted, the DVC status-checking logic would not be ableto keep track of it. This double-file design allows us to make subsequent steps depend on the actual results of the import, not our memory of the import in the Git repository.
Import steps are tracked in the
stage_status table in the database. For completed stages, this can
include a key (checksum, UUID, or other identifier) to identify a 'version' of the stage. Stages
can also have dependencies, which are solely used for computing the status of a stage (all actual
dependency relationships are handled by DVC):
stage_depstracks stage-to-stage dependencies, to say that one stage used another as input.
stage_filetracks stage-to-file dependencies, to say that a stage used a file as input.
source_file table tracks input file checksums.
Projects using the book database can also use
stage_status to obtain data version information, to
see if they are up-to-date.
bookdata package contains Python utility code, and the
src directory contains a number
of utility modules for use in the Rust code. To the extent reasonable, we have tried to mirror
design patterns and function names.
The general import philosophy is that we import the data into a PostgreSQL table in a raw form, only doing those conversions necessary to be able to access it with PostgreSQL commands and functions. We then extract information from this raw form into other tables and materialized views to enable relational queries.
Each data set's import process follows the following steps:
import/. This may be multiple steps; for example, OpenLibrary has a separate import step for each file. Actual import is usually handled by Rust or Python code.
Data integration then happens after the data sets are indexed (mostly - a few indexing steps depend on the book clustering process).
If you want to add a new data set, there are a few steps:
Set up the initial raw database schema, with an SQL script and corresponding DVC file under
schemas/. This should hold the data in a form that matches as closely as practical the raw
form of the data, and should have minimal indexes and constraints. For a new schema
you create two files:
ds-schema.sql, containing the CREATE SCHEMA and CREATE TABLE statements. We use PostgreSQL schemas (namespaces) to separate the data from different sources to make the whole database more manageable. Look at existing schema definitions for examples.
ds-schema.dvc, the DVC file running
ds-schema.sql. This should contain a few things:
cmd: python ../run.py sql-script ds-schema.sql
path: pgstat://common-schema outs:
path: pgstat://ds-schema cache: false ```
When you run
./dvc.sh repro schemas/ds-schema.dvc, it will run the schema script and
fill in the other values (e.g. checksums) for the dependencies and outputs.
Download the raw data files into
data and register them with DVC (
dvc add data/file for
the simplest case), and document in this file where to download them. For files that it is
reasonable to auto-download, you can create a more sophisticated setup to download them, but
this is often not necessary.
Identify, modify, and/or create the code needed to import the raw data files into the database. We have importers for several types of files already:
If the data is in CSV or similar form, suitable for PostgreSQL's
COPY FROM command, the
pcat import tool in the Rust tools can copy the file, decompressing if necessary, directly
to the database table.
If the data is in JSON, we have importers for two forms of JSON in the
tool, the source for which is in
src/commands/import_json/. Right now it supports
OpenLibrary and GoodReads JSON files; the first is a tab-separated file containing
object metadata and the JSON object, and the second is a simple object-per-line format.
The accompanying file (
goodreads.rs) define the data format and the
destination tables. For many future JSON objects,
goodreads.rs will be the appropriate
template to start with, and add support for it to the appropriate places in
If the data is in MARC-XML, the Rust
parse-marc command is your starting place. It can
process both multiple-record formats (e.g. from VIAF) or single-document formats (from the
Library of Congress), and can decompress while importing.
If you need to write new import code, you may need to make sure it properly records stage
dependencies and status. At a minimum, it should record each file imported and its checksum as
a file for the stage, along with the stage begin/end timestamps. Look at the
file for the specific tables. The
tracking.rs support modules provide code
for recording stage status in Python and Rust, respectively.
Set up the import process with an appropriate
.dvc step in
import/. This step should depend
on the schema (
pgstat://ds-schema), and have as one of its uncached outputs the import process
pgstat://ds-import, if the file is named
ds-import.dvc). Some importers require you
to explicitly provide the stage name as a command-line argument.
Write SQL commands to transform and index the imported data in a script under
index/. This script
may do a number of things:
See the existing index logic in
index/ for examples.
.dvc stage to run your index script; this works like the one for the schema in
(1), but is under
index/ and depends on
pgstat://ds-import (or whatever your import
stage is named).
Create or update data integrations to make use of this data, as needed and appropriate.
If the new data contains ISBN/ID links that you want to include in book clustering, add support
cluster.py and update the
cluster.dvc file to also depend on your data set's index
If appropriate, add a dependency on the last stage of your processing to
All dependencies should be through the
pgstat:// URLs, so that they are computed from current
Each data set comes with its own identification scheme for books:
We integrate these through two steps. First, we map ISBNs to numeric IDs with the
This table contains every ISBN (or ISBN-like thing, such as ASIN) we have seen and associates it
with a unique identifier.
Second, we map ISBN IDs to clusters with the
isbn_cluster table. A cluster is a collection of
related ISBNs, such as the different editions of a work. They correspond to GoodReads or OpenLibrary
'works' (in fact, when a GoodReads or OpenLibrary work is available, it is used to generate the
This allows us to connect ratings to metadata with maximal link coverage, by pulling in metadata across the whole book cluster.