Thoth Storages

GitHub tag (latest by date) PyPI - Module Version PyPI - License PyPI - Downloads

This library provides a library called thoth-storages used in project Thoth. The library exposes core queries and methods for PostgreSQL database as well as adapters for manipulating with Ceph via its S3 compatible API.

Installation and Usage

The library can be installed via pip or Pipenv from PyPI:

pipenv install thoth-storages

The library provides a CLI that can assist you with exploring schema and data storing:

thoth-storages --help
# In a cloned repo, run:
PYTHONPATH=. pipenv run python3 thoth-storages --help

You can run prepared test-suite via the following command:

pipenv install --dev
pipenv run python3 setup.py test

# To generate docs:
pipenv run python3 setup.py build_sphinx

Running PostgreSQL locally

You can use docker-compose.yaml present in this repository to run a local PostgreSQL instance, (make sure you installed podman-compose):

$ dnf install -y podman podman-compose
$ # Also available from PyPI: pip install podman-compose
$ podman-compose up

After running the commands above, you should be able to access a local PostgreSQL instance at localhost:5432. This is also the default configuration for PostgreSQL’s adapter that connects to localhost unless KNOWLEDGE_GRAPH_HOST is supplied explicitly (see also other environment variables in the adapter constructor for more info on configuring the connection). The default configuration uses database named postgres which can be accessed using postgres user and postgres password (SSL is disabled).

The provided docker-compose.yaml has also PGweb enabled to enable data exploration using UI. To access it visit localhost:8081.

The provided docker-compose.yaml does not use any volume. After you containers restart, the content will not be available anymore.

If you would like to experiment with PostgreSQL programmatically, you can use the following code snippet as a starting point:

from thoth.storages import GraphDatabase

graph = GraphDatabase()
graph.connect()
# To clear database:
# graph.drop_all()
# To initialize schema in the graph database:
# graph.initialize_schema()

Generating migrations and schema adjustment in deployment

If you make any changes to data model of the main PostgreSQL database, you need to generate migrations. These migrations state how to adjust already existing database with data in deployments. For this purpose, Alembic migrations are used. Alembic can (partially) automatically detect what has changed and how to adjust already existing database in a deployment.

Alembic uses incremental version control, where each migration is versioned and states how to migrate from previous state of database to the desired next state - these versions are present in alembic/versions directory and are automatically generated with procedure described bellow.

If you make any changes, follow the following steps which will generate version for you:

  • Make sure your local PostgreSQL instance is running (follow Running PostgreSQL locally instructions above):

    $ podman-compose up
    
  • Run Alembic CLI to generate versions for you:

    # Make sure you have your environment setup:
    # pipenv install --dev
    # Make sure you are running the most recent version of schema:
    $ PYTHONPATH=. pipenv run alembic upgrade head
    # Actually generate a new version:
    $ PYTHONPATH=. pipenv run alembic revision --autogenerate -m "Added row to calculate sum of sums which will be divided by 42"
    
  • Review migrations generated by Alembic. Note NOT all changes are automatically detected by Alembic.

  • Make sure generated migrations are part of your pull request so changes are propagated to deployments:

    $ git add thoth/storages/data/alembic/versions/
    
  • In a deployment, use Management API and its /graph/initialize endpoint to propagate database schema changes in deployment (Management API has to have recent schema changes present which are populated with new thoth-storages releases).

  • If running locally and you would like to propagate changes, run the following Alembic command to update migrations to the latest version:

    $ PYTHONPATH=. pipenv run alembic upgrade head
    

    If you would like to update schema programmatically run the following Python code:

    from thoth.storages import GraphDatabase
    
    graph = GraphDatabase()
    graph.connect()
    graph.initilize_schema()
    

When updating a deployment, make sure all the components use the same database schema. Metrics exposed from a deployment should state schema version of all the components in a deployment.

Generate schema images

You can use shipped CLI thoth-storages to automatically generate schema images out of the current models:

# First, make sure you have dev packages installed:
$ pipenv install --dev
$ PYTHONPATH=. pipenv run python3 ./thoth-storages generate-schema

The command above will produce an image named schema.png. Check --help to get more info on available options.

If the command above fails with the following exception:

FileNotFoundError: [Errno 2] "dot" not found in path.

make sure you have graphviz package installed:

dnf install -y graphviz

Creating own performance indicators

Performance indicators report performance aspect of a library on Amun and results can be automatically synced if the following procedure is respected.

To create own performance indicator, create a script which tests desired functionality of a library. An example can be matrix multiplication script present in thoth-station/performance repository. This script can be supplied to Dependency Monkey to validate certain combination of libraries in desired runtime and buildtime environment. Please follow instructions on how to create a performance script shown in the README of performance repo.

To create relevant models, adjust thoth/storages/graph/models_performance.py file and add your model. Describe parameters (reported in @parameters section of performance indicator result) and result (reported in @result). The name of class should match name which is reported by performance indicator run.

class PiMatmul(Base, BaseExtension, PerformanceIndicatorBase):
    """A class for representing a matrix multiplication micro-performance test."""

    # Device used during performance indicator run - CPU/GPU/TPU/...
    device = Column(String(128), nullable=False)
    matrix_size = Column(Integer, nullable=False)
    dtype = Column(String(128), nullable=False)
    reps = Column(Integer, nullable=False)
    elapsed = Column(Float, nullable=False)
    rate = Column(Float, nullable=False)

All the models use SQLAchemy. See docs for more info.

Online debugging of queries

You can print to logger all the queries that are performed to a PostgreSQL instance. To do so, set the following environment variable:

export THOTH_STORAGES_DEBUG_QUERIES=1

Memory usage statisticts

You can print information about PostgreSQL adapter together with statistics on the adapter in-memory cache usage to logger (it has to have at least level INFO set). To do so, set the following environment variable:

export THOTH_STORAGES_LOG_STATS=1

These statistics will be printed once the database adapter is destructed.

Automatic backups of Thoth deployment

In each deployment, an automatic knowledge graph backup cronjob is run, usually once a day. Results of automatic backups are stored on Ceph - you can find them in s3://<bucket-name>/<prefix>/<deployment-name>/graph-backup/pg_dump-<timestamp>.sql. Refer to deployment configuration for expansion of parameters in the path.

To create a database instance out of this backup file, run a fresh local PostgreSQL instance and fill it from the backup file:

$ cd thoth-station/storages
$ aws s3 --endpoint <ceph-s3-endpoint> cp s3://<bucket-name>/<prefix>/<deployment-name>/graph-backup/pg_dump-<timestamp> pg_dump-<timestamp>.sql
$ podman-compose up
$ psql -h localhost -p 5432 --username=postgres < pg_dump-<timestamp>.sql
password: <type password "postgres" here>
<logs will show up>

Manual backups of Thoth deployment

You can use pg_dump and psql utilities to create dumps and restore the database content from dumps. This tool is pre-installed in the container image which is running PostgreSQL so the only thing you need to do is execute pg_dump in Thoth’s deployment in a PostgreSQL container to create a dump, use oc cp to retrieve dump (or directly use oc exec and create the dump from the cluster) and subsequently psql to restore the database content. The prerequisite for this is to have access to the running container (edit rights).

# Execute the following commands from the root of this Git repo:
# List PostgreSQL pods running:
$ oc get pod -l name=postgresql
NAME                 READY     STATUS    RESTARTS   AGE
postgresql-1-glwnr   1/1       Running   0          3d
# Open remote shell to the running container in the PostgreSQL pod:
$ oc rsh -t postgresql-1-glwnr bash
# Perform dump of the database:
(cluster-postgres) $ pg_dump > pg_dump-$(date +"%s").sql
(cluster-postgres) $ ls pg_dump-*.sql   # Remember the current dump name
(cluster-postgres) pg_dump-1569491024.sql
(cluster-postgres) $ exit
# Copy the dump to the current dir:
$ oc cp thoth-test-core/postgresql-1-glwnr:/opt/app-root/src/pg_dump-1569491024.sql  .
# Start local PostgreSQL instance:
$ podman-compose up --detach
<logs will show up>
$ psql -h localhost -p 5432 --username=postgres < pg_dump-1569491024.sql
password: <type password "postgres" here>
<logs will show up>

You can ignore error messages related to an owner error like this:

STATEMENT:  ALTER TABLE public.python_software_stack OWNER TO thoth;
ERROR:  role "thoth" does not exist

The PostgreSQL container uses user “postgres” by default which is different from the one run in the cluster (“thoth”). The role assignment will simply not be created but data will be available.

Syncing results of a workflow run in the cluster

Each workflow task in the cluster reports a JSON which states necessary information about the task run (metadata) and actual results. These results of workflow tasks are stored on object storage Ceph via S3 compatible API and later on synced via graph syncs to the knowledge graph. The component responsible for graph syncs is graph-sync-job which is written generic enough to sync any data and report metrics about synced data so you don’t need to provide such logic on each new workload registered in the system. To sync your own results of job results (workload) done in the cluster, implement related syncing logic in the sync.py and register handler in the HANDLERS_MAPPING in the same file. The mapping maps prefix of the document id to the handler (function) which is responsible for syncing data into the knowledge base (please mind signatures of existing syncing functions to automatically integrate with sync_documents function which is called from graph-sync-job).

Query Naming conventions in Thoth

For query naming conventions, please read all the docs in conventions for query name.

Knowledge graph schema

Knowledge graph schema.

Crossroad

This documentation corresponds to implementation in version 0.57.3, documentation was generated on Oct 22, 2021.