Production and development¶
I use JASMIN first. Data-ocean is the development site and a copy of production data. A working page tells me the application is running; I still check the observation times before saying the data are current.
| Surface | Address | Purpose |
|---|---|---|
| Production dashboard | https://data.gamb2le.co.uk/app |
Normal operational browser view. |
| Production mobile API | https://data.gamb2le.co.uk/mobile/v1 |
Default for production iOS and Android apps. |
| Development dashboard/API | https://data-ocean.gamb2le.co.uk/app and /mobile/v1 |
Development views, including isolated candidate products. |
| Operational mobile mirror | https://data-ocean.gamb2le.co.uk/mobile/operational/v1 |
Dedicated read-only fallback of production products, when commissioned. |
The operational mirror is separately enabled. Its presence in this guide does not establish that it is deployed or fresh. The dated recovery record holds that evidence; see Recovery records.
Who owns the data¶
JASMIN aurora-cloud owns normal source syncs, product generation, Operations,
alerts, GWS/object-store archives and retention coordination. Production uses
aurora_site_env: production; development uses development, with normal
writer timers disabled. aurora_failover_role remains a compatibility setting.
Data-ocean aurora-cloud-droplet pulls production data in separate
aurora-dev-live-pull-<stage>.timer jobs. Most run every five minutes; AURORACam
and BLE stages run every two minutes. One slow stream must not hold the others.
Its ordinary writers are confined to /project/aurora/dev-raw and
/data/aurora/dev-products. Model evaluation is owned by its separate repository.
See Development experiments.
Raw observations land in /project/aurora/raw; products are in
/data/aurora/products. These paths are mirrored. The mirror is a continuity
copy, never archive parity or permission to prune field data.
Mirror and recovery rules¶
Every transfer checks the source directory and the source/destination mount.
Raw observations, service state and outage-only files are additive. Inventory
starts with deletion off and reconciliation on. After reviewed dry runs and
outage reconciliation, deletion can be enabled only for explicitly named
product-dashboard or product-quicklooks replicas. Their mounted source must
also be nonempty. Raw, Power, internal and state paths cannot opt into deletion.
For an existing host, playbooks/mirror_transport_release.yml replaces only
the transport used by the installed stage timers. Preserve the old wrapper and
supply its SHA-256 as aurora_mirror_transport_expected_sha256; review with
--check --diff first. The playbook requires additive reconciliation settings,
the existing planning-cache bind and its exact reviewed guard. It does not
alter schedules or credentials. A process already running finishes with its
old wrapper; verify the next scheduled stages before calling this deployed.
--partial and --delay-updates reduce exposure to incomplete files; they do
not make a whole multi-file Zarr transfer atomic. BLE holds its producer lock.
Always validate complete products and observation times before routing clients.
Keep existing development experiments and the relocated Power cache guard intact.
Each stage records its own success under
/var/lib/aurora-cloud/dev-live-mirror. The public
/data/aurora/internal/dev-live-mirror/last_success.json describes the dashboard
stage, not every stream. Compare source, raw, product, API and media frontiers.
What the apps do¶
Production apps start with JASMIN. A failed connection or HTTP 5xx response may use the dedicated operational mirror, labelled Development mirror with its data age. Authentication, certificate and schema failures do not trigger this fallback. Custom servers do not silently redirect to public infrastructure. The ordinary development API is not an automatic fallback: it may present candidate Power output. Related responses and media stay on one deployment; cached data remain labelled cached. Normal refresh tries JASMIN again.
Automatic client fallback does not move writer, archive, alert or retention ownership. Failover covers an actual ownership transfer.
Release policy¶
Completed work belongs on main; hosts run exact reviewed pins. Start with
Releases and use an isolated presentation checkout when source
integration includes scientific changes that have not been promoted.
I accept recovery after two successful scheduled cycles for each active stream, advancing observation times, explained intentional gaps and completed archive verification. Storage must cover existing service reserves and seven days of measured net growth. If that is not possible while preserving the observations, I hold the cutover and record the capacity needed. Retention stays paused until its existing verification gates pass.