QuotaClimat x Data For Good -

The aim of this work is to deliver a tool to a consortium around QuotaClimat, Climat Medias allowing them to quantify the media coverage of the climate crisis.

Radio and TV data are collected thanks to Mediatree API.

And webpress is currently at work in progress (as for 04/2024)

• 2022-09-28, Introduction by Eva Morel (Quota Climat): from 14:10 to 32:00 https://www.youtube.com/watch?v=GMrwDjq3rYs
• 2022-11-29 Project status and prospects by Estelle Rambier (Data): from 09:00 to 25:00 https://www.youtube.com/watch?v=cLGQxHJWwYA
• 2024-03 Project tech presentation by Paul Leclercq (Data) : https://www.youtube.com/watch?v=zWk4WLVC5Hs

Index

• I want to contribute! Where do I start?
• Development
  • File Structure
  • Setting up the environment

🤱 I want to contribute! Where do I start?

1. Learn about the project by watching the introduction videos mentioned above.
2. Create an issue or/and join https://dataforgood.fr/join and the Slack #offseason_quotaclimat.
3. Introduce yourself on Slack #offseason_quotaclimat

🔧 Development

Contributing

🔩 Setting up the environment

Doing the following step will enable your local environement to be aligned with the one of any other collaborator.

First install pyenv:

OS	Command
MacOS	cd - brew install pyenv # pyenv itself brew install pyenv-virtualenv # integration with Python virtualenvsec
Ubuntu	sudo apt-get update; sudo apt-get install make build-essential libssl-dev zlib1g-dev \ libbz2-dev libreadline-dev libsqlite3-dev wget curl llvm \ libncursesw5-dev xz-utils tk-dev libxml2-dev libxmlsec1-dev libffi-dev liblzma-dev curl https://pyenv.run | bash
Windows	An installation using miniconda is generally simpler than a pyenv one on Windows.

Make the shell pyenv aware:

OS	Command
MacOS	eval "$(pyenv init --path)" eval "$(pyenv init -)" eval "$(pyenv virtualenv-init -)"
Ubuntu	export PYENV_ROOT="$HOME/.pyenv" command -v pyenv >/dev/null || export PATH="$PYENV_ROOT/bin:$PATH" eval "$(pyenv init -)" eval "$(pyenv virtualenv-init -)"
Windows	🇫🇷 Dans Propriétés systèmes > Paramètres système avancés > Variables d'environnement... Choisissez la variable "Path" > Modifier... et ajoutez le chemin de votre installation python, où se trouve le python.exe. (par défaut, C:\Users\username\AppData\Roaming\Python\Scripts\ ) 🇬🇧 In System Properties > Advanced > Environment Variables... Choose the variable "Path" > Edit... et add the path to your python's installation, where is located the pyhton.exe (by default, this should be at C:\Users\username\AppData\Roaming\Python\Scripts\ ) In the console, you can now try : poetry --version

Let's install a python version (for windows, this step have been done with miniconda):

pyenv install 3.11.6 # this will take time

Check if it works properly, this command:

pyenv versions

should return:

  system
  3.11.6

Then you are ready to create a virtual environment. Go in the project folder, and run:

  pyenv virtualenv 3.11.6 quotaclimat
  pyenv local quotaclimat

In case of a version upgrade you can perform this command to switch

eval "$(pyenv init --path)"
pyenv activate 3.11.6/envs/quotaclimat

You now need a tool to manage dependencies. Let's use poetry. On windows, if not already installed, you will need a VS installation.

Link : https://wiki.python.org/moin/WindowsCompilers#Microsoft_Visual_C.2B-.2B-_14.x_with_Visual_Studio_2022_.28x86.2C_x64.2C_ARM.2C_ARM64.29

pip install poetry
poetry update
poetry lock

NLDA : I have not been able to work with wordcloud on windows.

When you need to install a new dependency (use a new package, e.g. nltk), run

poetry add ntlk

Update dependencies

poetry self update

After commiting to the repo, other team members will be able to use the exact same environment you are using.

Docker

First, have docker and compose installed on your computer

Then to start the different services

## To run only one service, have a look to docker-compose.yml and pick one service :
docker compose up metabase
docker compose up ingest_to_db
docker compose up mediatree
docker compose up test

docker secrets

inside the "secrets" folder, you should have these 4 files, you can put dummy values or ask Quota Climat team for the real ones.

secrets: # https://docs.docker.com/compose/use-secrets/
  pwd_api:
    file: secrets/pwd_api.txt
  username_api:
    file: secrets/username_api.txt
  bucket:
    file: secrets/scw_bucket.txt
  bucket_secret:
    file: secrets/scw_bucket_secret.txt

If you add a new dependency, don't forget to rebuild

docker compose build test # or ingest_to_db, mediatree etc

Explore postgres data using Metabase - a BI tool

docker compose up metabase -d

Will give you access to Metabase to explore the SQL table sitemap table or keywords here : http://localhost:3000/

To connect to it you have use the variables used inside docker-compose.yml :

• password: password
• username: user
• db: barometre
• host : postgres_db

Production metabase

If we encounter a OOM error, we can set this env variable : JAVA_OPTS=-Xmx2g

Web Press - How to scrap

The scrapping of sitemap.xml is done using the library advertools.

A great way to discover sitemap.xml is to check robots.txt page of websites : https://www.midilibre.fr/robots.txt

What medias to parse ? This document is a good start.

Learn more about site maps here.

Scrap every sitemaps

By default, we use a env variable ENV to only parse from localhost. If you set this value to another thing that docker or dev, it will parse everything.

Test

Thanks to the nginx container, we can have a local server for sitemap :

• http://localhost:8000/sitemap_news_figaro_3.xml

docker compose up -d nginx # used to scrap sitemap locally - a figaro like website with only 3 news
# docker compose up test with entrypoint modified to sleep
# docker exec test bash
pytest -vv --log-level DEBUG test # "test" is the folder containing tests
# Only one test
pytest -vv --log-level DEBUG -k detect
# OR
docker compose up test # test is the container name running pytest test

Deploy

Every commit on the main branch will build an deploy to the Scaleway container registry a new image that will be deployed. Have a look to .github/deploy-main.yml.

Learn more here.

Monitoring

With Sentry, with env variable SENTRY_DSN.

Learn more here : https://docs.sentry.io/platforms/python/configuration/options/

Mediatree - Import data

Mediatree Documentation API : https://keywords.mediatree.fr/docs/

You must contact QuotaClimat team to 2 files with the API's username and password inside :

• secrets/pwd_api.txt
• secrets/username_api.txt

Otherwise, a mock api response is available at https://github.com/dataforgoodfr/quotaclimat/blob/main/test/sitemap/mediatree.json

You can check the API with

curl -X POST https://keywords.mediatree.fr/api/auth/token/ \
               -H "Content-Type: application/x-www-form-urlencoded" \
               -d "grant_type=password" \
               -d "username=USERNAME" \
               -d "password=PASSWORD"

curl -X GET "https://keywords.mediatree.fr/api/epg/?channel=tf1&start_gte=2024-09-01T00:00:00&start_lte=2024-09-01T23:59:59&token=TOKEN_RECEIVED_FROM_PREVIOUS_QUERY"

Run

docker compose up mediatree

Configuration - Batch import

Based on time

If our media perimeter evolves, we have to reimport it all using env variable START_DATE like in docker compose (epoch second format : 1705409797). By default, it will import 1 day, you can modify it with NUMBER_OF_PREVIOUS_DAYS (integer).

Otherwise, default is yesterday midnight date (default cron job).

Production safety nets

As Scaleway Serverless service can be down, if some dates are missing until today, it will start back from the latest date saved until today.

Replay data

When dictionary change, we have to replay our data to update already saved data. As pandas to_sql with a little tweak can use upsert (update/insert), if we want to update already saved rows, we have to use :

• START_DATE
• NUMBER_OF_PREVIOUS_DAYS

For example to replay data from 2024-05-30 to 2024-05-01 we do from docker compose job "mediatree" (or scaleway job):

• START_DATE with unix timestamp of 2024-05-30 (1717020556)
• NUMBER_OF_PREVIOUS_DAYS to 30 to get back to 2024-05-01.

Warning: it might take several hours.

Based on channel

Use env variable CHANNEL like in docker compose (string: tf1)

Otherwise, default is all channels

Update without querying Mediatre API

In case we have a new word detection logic - and already saved data from Mediatree inside our DB (otherwise see Batch import based on time or channel) - we can re-apply it to all saved keywords inside our database.

⚠️ in this case, as we won't requery Mediatree API so we can miss some chunks, but it's faster. Choose wisely between importing/updating.

We should use env variable UPDATE like in docker compose (should be set to "true")

In order to see actual change in the local DB, run the test first docker compose up test and then these commands :

docker exec -ti quotaclimat-postgres_db-1 bash # or docker compose exec postgres_db bash
psql -h localhost --port 5432 -d barometre -U user
--> enter password : password
UPDATE keywords set number_of_keywords=1000 WHERE id = '71b8126a50c1ed2e5cb1eab00e4481c33587db478472c2c0e74325abb872bef6';
UPDATE keywords set number_of_keywords=1000 WHERE id = '975b41e76d298711cf55113a282e7f11c28157d761233838bb700253d47be262';

After having updated UPDATE env variable to true inside docker-compose.yml and running docker compose up mediatree you should see these logs :

 update_pg_keywords.py:20 | Difference old 1000 - new_number_of_keywords 0

We can adjust batch update with these env variables (as in the docker-compose.yml):

BATCH_SIZE: 50000 # number of records to update in one batch

Update only one channel

Use env variable CHANNEL like in docker compose (string: tf1) with UPDATE to true

Batch program data

UPDATE_PROGRAM_ONLY to true will only update program metadata, otherwise, it will update program metadata and all theme/keywords calculations.

UPDATE_PROGRAM_CHANNEL_EMPTY_ONLY to true will only update program metadata with empty value : "".

Batch update from a date

With +1 millions rows, we can update from an offset to fix a custom logic by using START_DATE_UPDATE (YYYY-MM-DD - default first day of the current month), the default will use the end of the month otherwise you can specify END_DATE (optional) (YYYY-MM-DD) to batch update PG from a date range.

Env variables list :

• START_DATE_UPDATE : string (YYYY-MM-DD ) - default to today - minus NUMBER_OF_DAYS (date is included in the query)
• END_DATE : string (YYYY-MM-DD ) - default to end of the month (date is included in the query)
• NUMBER_OF_DAYS : integer default to 7 days - number of days to update from (START_DATE_UPDATE - NUMBER_OF_DAYS) until START_DATE_UPDATE if START_DATE_UPDATE is empty
• STOP_WORD_KEYWORD_ONLY: boolean, default to False. If true will only update rows whose plaintext match top stop words' keyword. It uses to speed up update.
• BIODIVERSITY_ONLY: boolean (default=false), if true will only update rows that have at least one number_of_biodiversity_* > 0

Example inside the docker-compose.yml mediatree service -> START_DATE_UPDATE: 2024-04-01 - default END_DATE will be 2024-04-30

We can use a Github actions to start multiple update operations with different date, set it using the matrix

Production executions

~55 minutes to update 50K rows on a mVCPU 2240 - 4Gb RAM on Scaleway. Every month has ~80K rows.

SQL Tables evolution

Using Alembic Auto Generating Migrations¶ we can add a new column inside models.py and it will automatically make the schema evolution :

# If changes have already been applied (on your feature vranch) and you have to recreate your alembic file by doing :
# 1. change to your main branch 
git  switch main
# 2. start test container (docker compose up testconsole -d / docker compose exec testconsole bash) and run "pytest -vv -k api" to rebuild the state of the DB (or drop table the table you want) - just let it run a few seconds.
# 3. rechange to your WIP branch 
git switch -
# 4. connect to the test container : docker compose up testconsole -d / docker compose exec testconsole bash
# 5. reapply the latest saved state : 
poetry run alembic stamp head
# 6. Save the new columns
poetry run alembic revision --autogenerate -m "Add new column test for table keywords"
# this should generate a file to commit inside "alembic/versions"
# 7. to apply it we need to run, from our container
poetry run alembic upgrade head

Inside our Dockerfile_api_import, we call this line

# to migrate SQL tables schema if needed
RUN alembic upgrade head

Channel metadata

In order to maintain channel perimeter (weekday, hours) up to date, we save the current version inside postgres/channel_metadata.json, if we modify this file the next deploy will update every lines of inside Postgresql table channel_metadata.

Keywords

Produce keywords list from Excel files

How to update quotaclimat/data_processing/mediatree/keyword/keyword.py from shared excel files ? Download files locally to "document-experts" from Google Drive (ask on Slack) then :

Macro category sheet must be downloaded as a TSV as Dictionnaire - OME.xlsx - Catégories Transversales.tsv.

# Be sure to have updated the folder "document-experts" before running it :
poetry run python3 transform_excel_to_json.py

Program Metadata table

The media perimeter is defined here : "quotaclimat/data_processing/mediatree/channel_program_data.py"

To evolve the media perimeter, we use program_grid_start and program_grid_end columns to version all evolutions.

To calculate the right total duration for each channel, after updating "quotaclimat/data_processing/mediatree/channel_program_data.py" you need to execute this command to update postgres/program_metadata.json

poetry run python3 transform_program.py

The SQL queries are based on this file that generate the Program Metadata table.

Program data will not be updated to avoid lock concurrent issues when using UPDATE=true for keywords logic. Note: The default case will update them.

With the docker-entrypoint.sh this command is done automatically, so for production uses, you will not have to run this command.

Mediatre to S3

For a security nets, we have configured at data pipeline from Mediatree API to S3 (Object Storage Scaleway) with partition :

• country/year/month/day/channel If France, country code is None for legacy purposes.

Env variable used :

• START_DATE (integer) (unixtimestamp such as mediatree service)
• NUMBER_OF_PREVIOUS_DAYS (integer): default 7 days to check if something missing
• CHANNEL: (such as mediatree service)
• BUCKET : Scaleway Access key
• BUCKET_SECRET : Scaleway Secret key
• BUCKET_NAME
• DEFAULT_WINDOW_DURATION: int (default=20), the time window to divide the mediatree's 2 minute chunk (must be 120 secondes / DEFAULT_WINDOW_DURATION == 0)
• COUNTRY : 3 letter country code (default = fra - Source), see country.py to see them all - to get all countries the code is "all".

Stop words

To prevent advertising keywords to blow up statistics, we remove stop words based on the number of times a keyword is said in the same context.

The result will be saved inside postgresql table: stop_word.

This table is read by the service "mediatree" to remove stop words from the field "plaintext" to avoid to count them.

Env variables used :

• START_DATE (integer) (unixtimestamp such as mediatree service)
• NUMBER_OF_PREVIOUS_DAYS (integer): default 7 days
• MIN_REPETITION (integer) : default 15 - Number of minimum repetition of a stop word
• CONTEXT_TOTAL_LENGTH (integer) : default 80 - the length of the advertising context (sentence) saved
• FILTER_DAYS_STOP_WORD (integer): default 30 - number of days to filter the last stop words saved from - to speed up update execution

Remove a stop word

To remove a false positive, we set to false the validated attribute :

docker exec -ti quotaclimat-postgres_db-1 bash # or docker compose exec postgres_db bash
psql -h localhost --port 5432 -d barometre -U user
--> enter password : password
UPDATE stop_word set validated=false WHERE id = 'MY_ID';

Production monitoring

• Use scaleway
• Use [Ray dashboard] on port 8265

Bump version

poetry bump

poetry version minor

Materialized view - dbt

We can define some slow queries to make them efficient with materialized views using DBT, used via docker :

docker compose up testconsole -d
docker compose exec testconsole bash
> dbt debug  # check if this works
# caution: this seed will reinit the keywords and program_metadata tables
> dbt seed --select program_metadata --select keywords --full-refresh  # will empty your local db - order is important
> dbt run --models homepage_environment_by_media_by_month # change by your file name
> poetry run pytest --log-level DEBUG -vv my_dbt_project/pytest_tests # unit test 

Protips: Explore these data with postgres data using Metabase locally

DBT production

To update monthly our materialized view in production we have to use this command (automatically done inside our docker-entrypoint) that is run on every deployement of api-import (daily) :

poetry run dbt run --full-refresh

Causal query - too slow

Because this query is too massive, we set it month by month and avoid using a full-refresh. See units tests and docker-entrypoint.sh to see how.

If we change the DBT code, we have to relaunch this command to have a refreshed view (or wait the next daily cron).

SRT to Mediatree Format

Some Speech to Text data come from other sources than Mediatree, so we have to transform those source into the mediatree format to process them.

Run germany

Only for german data using parquet

docker compose up srt

or

docker compose up testconsole -d
docker compose exec testconsole bash
/app/ cd i8n/
/app/i8n# poetry run python3 srt-to-mediatree-format-parquet.py

Run belgium

Only for belgian data using .csv

Warning: this job is not automated as the process depending on getting the data is manual (emails), so we have to modify the script here.

docker compose up testconsole -d
docker compose exec testconsole bash
/app/ cd i8n/
/app/i8n# poetry run python3 srt-to-mediatree-format.py

Fix linting

Before committing, make sure that the line of codes you wrote are conform to PEP8 standard by running:

poetry run black .
poetry run isort .
poetry run flake8 .

There is a debt regarding the cleanest of the code right now. Let's just not make it worth for now.

Labelstudio

For the Climate Safeguards project we ingest the data present in the Labelstudio databases into the Barometre database. This way we can analyse the annotations of the factcheckers and extract key insights. There are two main tables that we need to ingest:

• task
• task_completion The task table contains an item (a 2 minute segment) with some metadata and an ID, and the task_completion table contains the annotations for that task. As the Labelstudio databases are separate from each other, we create a labelstudio_task_aggregate table and labelstudio_task_completion_aggregate table to perform a union of all the tasks and annotations. We also create task_aggregate_id and task_completion_aggregate_id to uniquely identify each task and annotation based on a hash of the table's id, project_id column and country column.

SQLAlchemy models

The models for the two tables can be found in quotaclimat/data_ingestion/labelstudio/models.py, any update to these models can be tracked with alembic, as the migration tool has been setup to track TargetBase as well as the already existing tables.

Source configuration

The sources for the Labelstudio ingestion can be found in quotaclimat/data_ingestion/labelstudio/configs.py. The db_config variable consists of a list of record (python dictionaries) with the source database name and a mapping of the project ids to countries:

db_config = [
  {
    "database": "<labelstudio_db>", 
    "countries": {
      1: "<country_1>", 
      2: "<country_2>", 
      3: "<country_3>",
    }
  },
]

When a new source is added (if a new Labelstudio instance is deployed) it suffices to add the source to the record list. (This assumes that all sources are on the same DB instace, as is the case at the time of writing).

Local execution

In order to execute the ingestion script locally you will need either have a working labelstudio locally, or to connect to the remote labelstudio with a read-only user. Set your credentials in the docker-compose.yml file:

LABELSTUDIO_INGESTION_POSTGRES_USER: <user>
LABELSTUDIO_INGESTION_POSTGRES_PASSWORD: <password>

and run the script via the test console:

docker compose up testconsole -d 
docker compose exec testconsole bash
poetry run python -m quotaclimat.data_ingestion.labelstudio.ingest_labelstudio

Analytics

In order to improve the performance of the dashboards hosted on Metabase, intermediate tables are calculated using dbt in the analytics schema. These can be found in my_dbt_project/models/analytics. The idea is to add a second layer to our database where we will store the more elaborated data used for our visualization. A schema of this evolution is seen below: These dbt models need to be run using the --target analytics command. You can test these locally using the test console:

docker compose up testconsole -d 
docker compose exec testconsole bash
# Seed the labelstudio tables
poetry run dbt seed --select program_metadatalabelstudio_task_aggregate --select labelstudio_task_completion_aggregate
# run the dbt model on the analytics target
poetry run dbt run --target analytics --select task_global_completion

Thanks

• [Paul Leclercq] (https://www.epauler.fr/)
• Eleven-Strategy
• Kevin Tessier