Entity Fragmentation

Generating graph data is a difficult process. The size of the datasets we want to process using followthemoney (FtM) makes it impossible to incrementally build nodes and edges in memory like you would in NetworkX. Instead, we use a stream-based solution for constructing graph entities. That is why the toolkit supports entity fragments and aggregation.


To illustrate this problem, imagine a table with millions of rows that describes a set of people and the companies they control. Every company can have multiple directors, while each director might control multiple companies:

A123Brilliant Amazing Ltd.John SmithPP8278171979-02-16
A71882Goldfish Ltd.John SmithPP827817NULL
A123Brilliant Amazing Ltd.Jane DoePP19882991983-06-24

Database humpty-dumpty

When turning this data into FtM, we’d create three entities for each row: a schema:Company, a schema:Person and a schema:Directorship that connects the two.

If we do this row by row, we’d eventually generate three schema:Company entities to represent two actual companies, and three schema:Person entities for two distinct people. Of course, we could write these to an ElasticSearch index sequentially - the later entities overwriting the earlier ones with the same ID.

That works only as long as each version of each entity contains the same data. In our example, the first mention of John Smith includes his birth date, while the second does not. If we don’t wish to lose that detail, we need to merge these fragments. While it’s possible to perform such merges at index time, this has proven to be impractically slow because it requires fetching each entity before it is updated.

A better solution is to sort all generated fragments before indexing them. With this approach, all the entities generated from the source table would be written to disk or to a database, and then sorted using their ID. In the resulting entity set, all instances of each company and person are grouped and can be merged as they are read.

In practice

In the FtM toolchain, there are two tools for doing entity aggregation: from the command-line ftm aggregate will merge fragments in memory. Alternately the add-on library followthemoney-store will perform the same operation in a SQLite or PostgreSQL database.

# Generate entities from a CSV file and a mapping:
cat company-registry.csv | ftm map-csv mapping_file.yml > fragments.ijson

# Write the fragments to a table `company_registry`:
cat fragments.ijson | ftm store write -d company_registry

# List the tables in the store:
ftm store list

# Output merged entities:
ftm store iterate -d company_registry

The same functionality can also be used as a Python library:

import os
from ftmstore import get_dataset
# Assume a function that will emit fragments:
from import generate_fragments

# If no `database_uri` is given, ftmstore will read connection from
# $FTM_STORE_URI, or create a file called `followthemoney.sqlite` in
# the current directory.
database_uri = os.environ.get('DATABASE_URI')
dataset = get_dataset('myapp_dataset', database_uri=database_uri)
bulk = dataset.bulk()
for idx, proxy in enumerate(generate_fragments()):
    bulk.put(proxy, fragment=idx)

# This will print the number of combined entities (ie. DISTINCT id):

# This will return combined entities:
for entity in dataset.iterate():

# You could also iterate the underlying fragments:
for proxy in dataset.partials():

# Note: `dataset.partials()` returns `EntityProxy` objects. The method
# `dataset.fragments()` would return raw Python dictionaries instead.

# All three methods also support the `entity_id` filter, which can also be
# shortened to `get`:
entity = dataset.get(entity_id)

Fragment origins

followthemoney-store is used across the tools built on FtM to capture and aggregate entity fragments. In Aleph, fragments for one entity might be written by different processes: the API, document ingestors, document NER analyzers or a translation backend. It is convenient to be able to flush all entity fragments from a particular origin, while leaving the other fragments intact. For example, this can be used to delete all data uploaded via the bulk API, while leaving document-based data in the same dataset intact.

To support this, ftm-store has the notion of an origin for each fragment. If specified, this can be used to later delete or overwrite subsets of fragments.

cat us_ofac.ijson | ftm store write -d sanctions -o us_ofac
cat eu_eeas.ijson | ftm store write -d sanctions -o eu_eeas

# Will now have entities from both source files:
ftm store iterate -d sanctions | wc -l

# Delete all fragments from the second file:
ftm store delete -d sanctions -o eu_eeas

# Only one source file is left:
ftm store iterate -d sanctions | wc -l