Urban Observations

Urban Observations collects live and daily data for Los Angeles from public feeds, authenticated APIs, traffic and wildfire cameras, and email alerts. Raw observations are written to a configurable data directory for use by SIGMUS, IncidentLens, or other analysis pipelines.

Requirements

• Enough storage for camera images and daily archives
• Accounts or API keys for PeMS, OpenWeather, PurpleAir, and Gmail
• Docker Engine with the Compose plugin
• Python 3 on the host (standard library only, used by docker-start to read JSON)

Configuration

The deployment reads operator settings and credentials from config.json. Create it by copying config.example.json; the file is ignored by Git because it contains plaintext secrets.

Field	What to enter	Used by
save_folder	Data directory on the host. Relative paths such as ./data start at the repository; absolute paths such as /media/user/External Drive/data are also supported.	All extractors and the data check
backup_folder	Archive directory on the host, with the same relative-or-absolute path rules as save_folder.	Cleanup and the data check
retention_days	Number of recent daily data directories to retain locally before cleanup archives older days. Use a positive integer, such as 30.	cleanup
pem_username	Caltrans PeMS account username.	pems
pem_password	Caltrans PeMS account password.	pems
open_weather_map_api_key	OpenWeather API key.	weather
purpleair_api_key	PurpleAir read API key.	air and the sensor-inventory refresh command
email_acc_info.email	Gmail address that receives X/IFTTT and Citizen messages and sends data-quality alerts.	twitter-email, citizen-email, and data-alert
email_acc_info.password	Gmail app password for that address, not its normal account password.	Email services and data-alert
data_quality_alerts.recipient	Address that should receive missing-data alerts; it can differ from the ingestion Gmail address.	data-alert

The checked-in weather locations, PurpleAir sensors, and CCTV document are collector resources rather than configuration fields. Instructions for changing the PurpleAir inventory appear later in this README.

Docker itself cannot read JSON values while constructing bind mounts. The included docker-start wrapper reads these two fields before invoking Compose, resolves and creates the directories, and mounts only those exact paths. This supports external disks and paths containing spaces without exposing all of /media, /mnt, or the host filesystem to every container. The wrapper does not schedule or supervise jobs; Docker Compose still does that work.

Quick start

Copy the example once, put your accounts and API keys in config.json, and start everything (by everything, we mean all services shown in the below table):

cp config.example.json config.json
chmod 600 config.json
./docker-start up

Compose builds eight dependency-focused images from shared layers and runs ten independently scheduled containers. Every image inherits the same slim Python base plus the scheduler and configuration loader; each target then adds only its collector code and dependencies. The two email services share the email image, and the two operational services share the operations image. Chromium is installed only in the ALERTCalifornia image.

Compose service	Image	Kind	Access mechanism and purpose
gdelt-gkg	urban-data-collector-gdelt	Extractor	Polls the public GDELT event/GKG HTTP feeds, saves Los Angeles matches, and downloads and parses the webpages referenced by matching GKG rows.
pems	urban-data-collector-pems	Extractor	Signs in to the Caltrans PeMS Clearinghouse and downloads District 7 traffic-station and CHP-incident files.
cctv	urban-data-collector-cctv	Extractor	Scrapes the Caltrans camera stream list and downloads snapshots from Los Angeles County cameras.
alertcalifornia	urban-data-collector-alertcalifornia	Extractor	Uses Chromium browser automation to discover ALERTCalifornia cameras and capture images and locations.
weather	urban-data-collector-weather	Extractor	Calls the OpenWeather API for every coordinate in the checked-in weather inventory.
air	urban-data-collector-air	Extractor	Calls the PurpleAir API for every sensor in the checked-in PurpleAir inventory.
twitter-email	urban-data-collector-email	Extractor	Reads new X/IFTTT notification messages from Gmail over IMAP and stores the raw messages.
citizen-email	urban-data-collector-email	Extractor	Reads new Citizen alert messages from Gmail over IMAP and stores the raw messages.
cleanup	urban-data-collector-operations	Operation	Archives completed data days to the backup directory and removes old local days only after archive validation.
data-alert	urban-data-collector-operations	Operation	Checks that expected daily datasets exist locally or in backup archives and sends a missing-data email over SMTP.

Alternative ways to run

If you want to just run a single service (not all of them), start one service or a selected group by putting its service name after up:

./docker-start up air
./docker-start up air weather cctv

There is no .env file to copy or maintain. config.json is ignored by Git and is the single credential/configuration file. It contains plaintext credentials, so keep its permissions restricted and never commit it.

Use the wrapper for lifecycle commands so Compose always receives the storage mounts selected by config.json:

./docker-start ps
./docker-start logs --follow cctv
./docker-start stop cctv
./docker-start down

Omit the service after logs -f to follow everything. With the example configuration, Compose stores collected data in ./data, backups in ./backups, and uses America/Los_Angeles for schedules. Change the two paths in config.json to choose other relative directories or absolute locations on the host, including mounted external drives.

Change collection frequency

Edit the service's --cron value in compose.yaml, then recreate that service. Schedules use five-field cron syntax: minute, hour, day of month, month, and day of week. For example, change the weather command from hourly:

"--cron", "0 * * * *"

to every 30 minutes:

"--cron", "*/30 * * * *"

Apply the change with ./docker-start up weather.

Each source is a long-running container with its own APScheduler process. A job is never run more than once concurrently within its service; failures are logged and the next scheduled run still occurs. Scheduled times use America/Los_Angeles by default.

The repository already contains the default PurpleAir sensor inventory, so no initialization command is required before starting the air service. See "PurpleAir sensor inventory" below when you want to replace that list.

After startup, confirm that the schedulers are running and inspect each source before relying on unattended collection:

./docker-start ps
./docker-start logs --tail=50 weather cctv alertcalifornia
find ./data -type f | head

The email-backed collectors preserve messages in the ingestion mailbox. Each stores a stable IMAP UID checkpoint beneath <save_folder>/.email_state, so it processes only mail that arrived after the previous successful run. Test email credentials using a dedicated mailbox before enabling those services. ./docker-start up enables all services; use an explicit service list for a staged rollout.

Do not start a second scheduler against the same mounted dataset because that would duplicate collection and could race on the email mailbox.

Configuration and credentials

Fill the corresponding values directly in config.json; use config.example.json as the complete template. Account settings include pem_username, pem_password, open_weather_map_api_key, purpleair_api_key, email_acc_info, and the alert recipient under data_quality_alerts.

Set the destination for missing-data alerts explicitly. It may be the ingestion mailbox or a different operations address:

"data_quality_alerts": {
  "recipient": "operations@example.com"
}

The Gmail password must be an app password, not the normal account password. The collectors preserve messages and use UID checkpoints to process only new mail. The alert recipient may be different from the ingestion email address.

The OpenWeather coordinate list, PurpleAir sensor inventory, and Caltrans CCTV document are application resources. Their stable paths are owned by the relevant collector/container and do not belong in config.json.

Provider account setup

PurpleAir

Create a developer account at PurpleAir Develop, create a read API key, and place it in config.json as purpleair_api_key. Never commit that file.

With the current 50-sensor hourly inventory, budget approximately 2,500 PurpleAir points per day. A fresh allocation of 1,000,000 free points lasts about 400 days, or roughly 13 months, at that rate. For the currently deployed account, set an operational reminder to replace or replenish the account/key in early May 2027; that date reflects the existing deployment's usage timeline, not 13 months from a newly created account.

Caltrans PeMS

Apply for a PeMS account, then place the approved username and password in pem_username and pem_password. PeMS indicates that account applications involve manual review, so allow time for a person to approve the account before testing the extractor.

OpenWeather

Create an OpenWeather account, generate an API key, and place it in open_weather_map_api_key. The free API currently advertises limits of 60 API calls per minute and 1,000,000 calls per month. The hourly collector makes one call per predefined coordinate, so keep the number of entries in the checked-in OpenWeather location inventory within the account quota. See the OpenWeather pricing and limits before a new deployment in case the provider changes them.

IFTTT, X, and Citizen

X and Citizen notification ingestion requires IFTTT Pro for the applets used by this deployment. Budget approximately $3.49 per month, but verify the current IFTTT plan price; pricing and annual-billing discounts can change.

An IFTTT applet's setup/status screen may continue to display the destination address captured when the applet was created (for example, an old testingemail_...@gmail.com address). In practice, the Email action sends to the Email field on the current IFTTT account. Check that account field when changing ingestion mailboxes. It can be different from the address originally used to sign up for IFTTT.

Gmail app password

The extractors connect to Gmail through IMAP and the alert monitor sends through SMTP. Enable 2-Step Verification on the Gmail account, then create a 16-digit Google app password and put it in email_acc_info.password. Do not use or store the normal Google account password. Google may revoke app passwords after the Google account password is changed, in which case generate a replacement and update config.json.

Data-quality alert defaults

Normal installations do not need a data_quality_alerts object in config.json. The monitor defaults to Gmail SMTP (smtp.gmail.com:465) and the following expectations:

• A value of 1 means that yesterday's folder/archive must exist.
• A value of 2 means that the folder/archive from two days ago must exist, giving the provider one extra publication day.
• Air, ALERTCalifornia, CCTV, GKG, and weather use 1.
• Both PeMS products use 2 because PeMS publication is delayed.
• Event-driven X/Citizen and disabled noise/seismic are not checked by default.

Set only data_quality_alerts.recipient for the standard deployment. Advanced users can override the defaults with an optional configuration object, for example:

{
  "data_quality_alerts": {
    "recipient": "operations@example.com",
    "sources": {"air_data": 1, "pem_data_station_5min": 2}
  }
}

When sources is supplied it replaces, rather than extends, the built-in source map. SMTP host and port should be overridden only when using a non-Gmail mail server.

Configure notification sources

X and Citizen data arrive through the Gmail account configured above.

X

Create an IFTTT applet using Email new tweets from a specific X user. Configure one applet for each X account to monitor and send the messages to the ingestion mailbox. The extractor recognizes notification senders or subjects containing twitter or x.com and stores the original message without interpreting it. Event and location extraction belongs to downstream ingestion.

Citizen

Register for the desired location alerts in the Citizen app and have those alerts delivered to the same ingestion mailbox. The Citizen extractor stores the original headers and body without parsing them. No Citizen API or OpenAI credential is required by this repository.

PurpleAir sensor inventory

PurpleAir observations require a local inventory of sensor IDs and locations. The tracked file extractor_modules/air/nearby_purpleair_sensors.csv is the inventory used by default. Because it is part of the repository and Docker image, a fresh checkout can collect from those sensors immediately. It uses this format:

sensor_index,latitude,longitude

The file has no header. Each row contains a numeric PurpleAir sensor index, latitude, and longitude.

To replace the inventory, provide a center point and radius. This command reads purpleair_api_key from the current config.json, queries PurpleAir for outdoor sensors, selects up to 50 sensors distributed across the search area, and overwrites the tracked CSV:

./docker-start compose run --rm air \
  python -m extractor_modules.air.refresh_sensors \
  --location 34.0549 -118.2426 \
  --radius-km 30 \
  --sensor-count 50

The coordinates above are downtown Los Angeles. --location takes latitude then longitude; --radius-km must be positive. Review the resulting Git diff before committing it:

git diff -- extractor_modules/air/nearby_purpleair_sensors.csv

This calls PurpleAir's Get Sensors API, selects outdoor sensors around the requested location and writes the same three-column format consumed by the hourly collector. API calls consume PurpleAir points. PurpleAir data is subject to its Data Licensing requirements, including attribution and distribution restrictions. Confirm that committing and distributing an updated inventory complies with the terms applicable to your PurpleAir account.

Data sources and collection schedule

compose.yaml contains the production schedules.

Source	Data access	Frequency	Output beneath save_folder
GDELT Events/GKG	http://data.gdeltproject.org/gdeltv2/lastupdate.txt and the CSV ZIP URLs listed there	every 15 minutes	gkg/YYYYMMDD/
Caltrans PeMS	Authenticated Clearinghouse downloads from https://pems.dot.ca.gov; District 7 station_5min and chp_incidents_day files	daily at 08:00	pem_data_station_5min/YYYYMMDD/, pem_data_chp_incidents_day/YYYYMMDD/
Caltrans CCTV	Camera list and streams discovered from https://cwwp2.dot.ca.gov/vm/streamlist.htm; LA County cameras	every 15 minutes	cctv/YYYYMMDD/
ALERTCalifornia	Cameras discovered from https://cameras.alertcalifornia.org/?pos=33.9639_-118.2898_10; LA County bounds	every 30 minutes	alertcalifornia/YYYYMMDD/
OpenWeather	OpenWeather API at the locations in the checked-in weather inventory; includes wind speed and direction	hourly	weather_data/YYYYMMDD/
PurpleAir	PurpleAir API for the tracked sensor inventory in extractor_modules/air	hourly	air_data/YYYYMMDD/
X notifications	IFTTT email notifications read through imap.gmail.com	every 5 minutes	twitter_data/YYYYMMDD/
Citizen notifications	Registered Citizen alert emails read through imap.gmail.com	every 30 minutes	citizen_data/YYYYMMDD/
Cleanup and backup	Local storage maintenance	daily at 08:00	archives completed days under backup_folder
Data completeness alert	Local folders and backup archives; email sent with the ingestion Gmail account	daily at 10:00	no data output; set data_quality_alerts.recipient

NoiseCapture and SCEDC seismic collector code is included but has no service in compose.yaml, so it is disabled by default. Test it manually and add an explicit Compose service before enabling it.

The Los Angeles deployment also uses these checked-in inventories:

• extractor_modules/weather/owm_locations.txt
• extractor_modules/air/nearby_purpleair_sensors.csv
• extractor_modules/cctv/cctv.kml
• extractor_modules/seismic/seismic_stations.txt

Collection scope and operator choices

The collectors do not all discover devices in the same way. Some query a provider for everything in a geographic area, while others depend on an operator-maintained inventory or a deliberately selected subset.

Source	How collection scope is chosen	Manual choice or dynamic query
PurpleAir	The extractor_modules.air.refresh_sensors command queries outdoor sensors around the requested --location and --radius-km, then K-means selects the requested number of geographically distributed sensors and updates the tracked inventory. Hourly collection uses only that saved inventory.	Both. Geographic discovery is automatic, while the search area and sensor count are operator choices and cost controls.
OpenWeather	Calls the OpenWeather Current Weather API for the coordinates in extractor_modules/weather/owm_locations.txt.	Manual/predefined coordinate inventory. Add or remove locations by editing the inventory.
Caltrans CCTV	Queries the Caltrans stream-list page on every run and selects every camera whose county is Los Angeles; optional include/exclude lists can narrow it.	Dynamic geographic/administrative query, with optional manual filtering.
ALERTCalifornia	Opens the ALERTCalifornia gallery at the configured LA-area map bounds, discovers the cameras currently exposed by the site, and optionally applies include/exclude lists.	Dynamic geographic query, with the map bounds and optional filters chosen by the operator.
Caltrans PeMS	Logs into the Clearinghouse and downloads the newest District 7 files for the two configured file types.	Manual district and product selection; file discovery is dynamic.
GDELT GKG/events	Downloads the latest GDELT update and retains rows mentioning Los Angeles in the configured location fields.	Dynamic feed query with a hard-coded LA filter.
X and Citizen	Reads new messages delivered to the configured ingestion mailbox and advances a persistent IMAP UID checkpoint.	Event-driven. The mailbox, notification subscriptions, and subject/content rules define scope. Messages remain in the inbox.
Noise and seismic	Use their checked-in station/configuration files when enabled.	Manual inventories; disabled by default.

Current weather conditions are obtained from the OpenWeather Current Weather API at predefined geographic coordinates. OpenWeather produces these conditions using weather models and observational sources including weather stations, radar, and satellite data.

Data fields and file formats

Raw provider values are preserved wherever practical. Most CSV files have no header unless explicitly noted below, so consumers should use the documented column order rather than infer it from the first record.

Source	Stored fields/content
PurpleAir	Headerless CSV rows: sensor_index, latitude, longitude, pm2_5. Each hourly file contains one row per selected sensor.
OpenWeather	Headerless CSV row per location: temperature_f, detailed_status, humidity_percent, wind_speed_m_s, wind_direction_degrees. The location name is represented by its parent folder; coordinates come from the checked-in weather inventory. Wind direction is meteorological degrees reported by OpenWeather.
Caltrans CCTV	JPEG snapshot. Camera identity is the parent folder and capture time is the filename. County, nearby place, camera name, and image URL are used during discovery but are not repeated inside the JPEG.
ALERTCalifornia	JPEG snapshot plus a same-timestamp .location text file containing latitude,longitude,camera_direction. Camera identity is the parent folder.
PeMS station 5-minute	Provider .txt.gz file. Rows contain the Caltrans station-observation schema, including timestamp, station/district/freeway/direction/lane type and aggregate/per-lane traffic measurements such as flow, occupancy, and speed. The extractor does not rewrite provider columns.
PeMS CHP incidents	Provider .txt.zip daily incident product. Incident fields and encoding are retained exactly as published by the PeMS Clearinghouse.
GDELT GKG/events	CSV with the original provider columns retained and numbered (0, 1, ...). The extractor writes both filtered GKG and event/export products into gkg; rows are filtered for Los Angeles before storage. Each GKG CSV also has a same-stem directory (for example, 20260805070000.gkg/) containing SHA-256-named raw .html, parsed-body .txt, and request/parser .json files plus manifest.json. Parsing uses Newspaper without an LLM. Failed or blocked URLs retain JSON error metadata.
X notifications	Raw email_raw.v1 CSV: schema_version, source, imap_uid, message_id, received_at, sender, subject, body, ingested_at.
Citizen notifications	The same raw email_raw.v1 CSV contract. Files are written only when messages are processed.

Historical X and Citizen folders may contain the older enriched CSV schemas. Those archives are intentionally not rewritten. SIGMUS and IncidentLens readers accept both legacy enriched rows and the current raw schema.

Observed daily storage estimates

These are planning estimates measured from the Los Angeles deployment's local daily directories on 2026-08-31. They are uncompressed source-directory sizes, not guaranteed quotas. Partial days, unavailable cameras, news volume, and the number of incoming emails can change the result substantially.

Source	Observed daily size	Planning guidance
PurpleAir, 50 sensors hourly	about 30 KB	Scales approximately with sensor count and sampling frequency.
OpenWeather predefined locations hourly	about 45 KB	Scales with the number of configured coordinates.
Caltrans CCTV	about 0.5–0.65 GB on a full day	Largest source; the observed 30-day average was about 536 MB/day, including some partial days.
ALERTCalifornia	about 2–48 MB in the available sample	Highly dependent on site availability, discovered cameras, and successful browser runs.
PeMS station 5-minute	about 17–19 MB	One compressed District 7 daily file.
PeMS CHP incidents	about 0.3–0.45 MB	One compressed daily incident file.
GDELT GKG/events and article pages	Provider CSVs average about 13 MB/day. Article HTML, parsed text, and metadata add an estimated 0.4 GB/day at roughly 1,170 URLs/day (one live 47-URL interval used 17 MB).	News volume, page size, publisher blocking, and LA matches vary substantially; allow at least 0.5 GB/day.
X notifications	typically 5–20 KB on active days	Event-driven; days without matching messages may have no folder.
Citizen notifications	typically 4–10 KB in the current sample	Event-driven and dependent on message volume and raw email content.
Noise/seismic	insufficient current data for a reliable estimate	Disabled by default; measure after enabling with the intended inventory.

With the current inventories and GKG article capture, a normal day is expected to use roughly 0.95–1.25 GB before archive/tar overhead. Thirty local calendar days therefore require approximately 29–38 GB under present conditions. Capacity planning should include additional headroom for full camera days, unusually large news days, and future inventory growth.

The OpenWeather and PurpleAir inventories are part of their collectors. Some LA geographic bounds are defined in extractor code, so replacing an inventory alone does not fully retarget the system to another city.

The included extractor_modules/cctv/cctv.kml comes from Caltrans' California Open Data dataset Closed Circuit Television, which describes CCTV locations on the State Highway Network and is published under the Creative Commons Attribution license. Source: California Department of Transportation (Caltrans).

Run a service immediately for testing

docker-start run-now creates a temporary container with the selected service's configuration and exact storage mounts. It replaces the normal scheduler with that service's extractor, so collection starts immediately and the temporary container exits when finished. It does not wait for the configured cron frequency.

For example, run CCTV now:

./docker-start run-now cctv

The equivalent commands for all extractors are:

./docker-start run-now gdelt-gkg
./docker-start run-now pems
./docker-start run-now cctv
./docker-start run-now alertcalifornia
./docker-start run-now weather
./docker-start run-now air
./docker-start run-now twitter-email
./docker-start run-now citizen-email

The operational containers can also be tested without changing data or sending an alert:

./docker-start run-now cleanup --dry-run
./docker-start run-now data-alert --dry-run

These commands contact live services. PeMS, OpenWeather, PurpleAir, and email collection fail if their corresponding credentials are absent or invalid. Verify that a new dated directory and output file appear beneath the configured save_folder before relying on its scheduled service.

If the scheduled service is already running, its next run could overlap the temporary test container. For debugging a collector in isolation, stop and restart only that service around the test:

./docker-start stop cctv
./docker-start run-now cctv
./docker-start start cctv

Storage layout

<save_folder>/
├── .email_state/{twitter,citizen}.json
├── air_data/YYYYMMDD/YYYYMMDDHHMMSS.csv
├── alertcalifornia/YYYYMMDD/<camera>/
│   ├── YYYYMMDDHHMMSS.jpg
│   └── YYYYMMDDHHMMSS.location
├── citizen_data/YYYYMMDD/<epoch_ms>.csv
├── cctv/YYYYMMDD/<camera>/YYYYMMDDHHMMSS.jpg
├── gkg/YYYYMMDD/
│   ├── <provider_timestamp>.*.csv
│   └── <provider_timestamp>.gkg/               # article capture for each GKG interval
│       ├── <url_hash>.html
│       ├── <url_hash>.txt
│       ├── <url_hash>.json
│       └── manifest.json
├── noise_planet/YYYYMMDD/                       # if enabled
├── pem_data_chp_incidents_day/YYYYMMDD/*.txt.zip
├── pem_data_station_5min/YYYYMMDD/*.txt.gz
├── seismic/YYYYMMDD/<station>/                  # if enabled
├── twitter_data/YYYYMMDD/<epoch_ms>.csv
└── weather_data/YYYYMMDD/<location>/YYYYMMDDHHMMSS.csv

<backup_folder>/
└── raw/
    ├── air_data/YYYYMMDD.tar
    ├── cctv/YYYYMMDD.tar
    ├── weather_data/YYYYMMDD.tar
    └── <each-other-source>/YYYYMMDD.tar

The intended production configuration places backup_folder on the external drive. Every completed date for every source is pushed there as <backup_folder>/raw/<source>/YYYYMMDD.tar; archives are retained on the external drive indefinitely unless an operator applies a separate archive retention policy. The current day is deliberately not archived while collectors may still be writing it, and is pushed during the next daily cleanup.

The local disk retains retention_days calendar days, including today (30 by default). Once a local date is older than that window, it is deleted only after its external archive has been created and validated. Existing archives are refreshed when their source day changes. Today's directory and non-date entries such as .config and .email_state are never cleanup candidates. If the external drive is absent, read-only, or an archive cannot be validated, cleanup fails safely before deleting any unverified local data.

Before enabling automatic cleanup in either deployment mode, preview the plan and validate existing archives without changing data:

python -m extractor_modules.operations.archive --dry-run

Use --max-days N for a one-time retention override. During a real run, new archives are written to temporary files, validated, and atomically renamed. The deletion phase starts only after every completed day has a readable archive. If any archive fails, the command exits nonzero before deleting raw directories. On an existing installation, the first successful run may need substantial temporary and backup capacity; review the dry-run summary and available space before running without --dry-run.

The 10:00 completeness monitor checks both local date folders and archives, because cleanup may already have moved yesterday off local storage. Continuous feeds are expected through yesterday; PeMS receives one additional day of publication grace. Twitter, Citizen, noise, and seismic are excluded by default because they are event-driven or disabled. Alerts authenticate to Gmail using email_acc_info and are sent to data_quality_alerts.recipient.

Troubleshooting

• Check ./docker-start logs <service> and the modification time of each source's newest file.
• If ALERTCalifornia stops after a browser update, rebuild the image so Chromium and chromedriver are updated together. For a local installation, ensure their versions are compatible.
• GDELT lastupdate.txt can lag or point to an incomplete download.
• PeMS daily files may be published late, especially near month boundaries.
• If a container reports a missing credential, confirm the corresponding value is present directly in config.json.
• If files are written to an unexpected location, check save_folder and backup_folder in config.json. Relative paths start at the repository; absolute paths are used directly.
• Do not enable HTTP-level debugging in the PeMS client: request tracing can include the login form and expose credentials in container logs.

Tests

python -m pip install -r requirements/dev.txt
python -m pytest -q

Tests remain in the source repository for regression checking but are not copied into the production container image.