Data classification & retention
This policy sets out how data handled by the Edmund platform is classified, how long it is retained, and how it is disposed of at the end of its retention period. All data is categorized by its sensitivity, the impact of its compromise, and the regulatory requirements that apply to it.
Data classification
Public data
-
Definition: Data intentionally made available to the general public, whose disclosure would not cause harm to the organization or to individuals.
-
Examples: Public marketing materials, general product information, public FAQs.
-
Handling: No specific access restrictions. Minimal security controls required.
Internal data
-
Definition: Data intended for internal use only. Unauthorized disclosure could cause minor operational disruption or reputational damage.
-
Examples: Internal memos, non-sensitive project plans, general employee contact lists.
-
Handling: Restricted to authorized personnel within the organization. Standard security measures applied.
Confidential data
-
Definition: Data whose unauthorized disclosure could lead to significant financial loss, legal penalties, reputational damage, or a privacy breach. This includes Personally Identifiable Information (PII) and sensitive business information.
-
Examples: Customer PII (names, addresses, e-mail addresses, phone numbers), financial records, proprietary algorithms, trade secrets.
-
Handling: Strict access controls, encryption at rest and in transit, regular auditing, and adherence to privacy regulations such as the GDPR. Access is granted on a need-to-know basis.
Restricted data
-
Definition: The most sensitive category, requiring the highest level of protection. Unauthorized disclosure would have severe consequences, including regulatory penalties, major financial loss, and irreparable damage to reputation.
-
Examples: Production access credentials and secure access keys, highly sensitive intellectual property.
-
Handling: Extremely stringent access controls, multi-factor authentication, advanced encryption, and continuous monitoring. Access is highly restricted and requires explicit authorization.
Retention periods
Data is retained for a period determined by its classification, the applicable legal and regulatory requirements, business needs, and historical value.
| Classification | Retention period | Disposal |
|---|---|---|
| Public | Retained indefinitely, or until no longer relevant for public access. | Removed when outdated or superseded. |
| Internal | Typically 1–3 years, or as long as necessary for operational purposes. | Reviewed regularly and securely deleted or archived when no longer needed. |
| Confidential | Varies by data type and regulatory requirement — for example, records kept 7 years for tax purposes, and PII for as long as necessary to provide the service plus a grace period. | Securely deleted, anonymized or pseudonymized after the retention period, in compliance with applicable law. |
| Restricted | Defined by the mandate that applies to the specific data, and kept no longer than that mandate requires. | Secure and auditable destruction, with a record of the disposal retained. |
Data disposal
All data, regardless of classification, is securely disposed of at the end of its retention period. Disposal methods ensure that the data cannot be reconstructed or accessed:
-
Digital data: Secure wiping, degaussing, or physical destruction of the storage media.
-
Physical data: Shredding or incineration of documents.
Sub-processors and retention
Every sub-processor Edmund uses carries a Data Processing Agreement and operates within the European Union. Customer document content and application data remain in the customer’s database; see Architecture & data flow.
| Sub-processor | Purpose | Region | Retention |
|---|---|---|---|
| Amazon Web Services (EMEA SARL) | Platform hosting | Ireland, eu-west-1 | For the duration of the service |
| Zilliz Cloud (Milvus) | Vector database | Germany, eu-central-1 | For the duration of the service |
| Azure OpenAI (Microsoft Foundry) | LLM provider | Sweden, swedencentral | Up to 30 days, for abuse monitoring |
| Microsoft Azure | Telemetry storage | Netherlands, westeurope | 14 days (application-level) |
| Sentry | Error monitoring | Germany, de.sentry.io | 90 days |
| LangSmith | LLM tracing, debugging | Netherlands, europe-west | Up to 400 days |
Azure OpenAI does not train its models on customer data. For abuse monitoring the provider retains a temporary sample of prompts for no more than 30 days, used solely to detect misuse of the service.
Sentry is configured to filter or redact sensitive information — passwords, tokens and PII — from error reports and stack traces. Collection of personal data is disabled in the client configuration, and the language-model integration is switched off, so prompts and model responses are never transmitted to Sentry. Only the contextual data required to debug an error is sent.
LangSmith receives the prompt sent to the language model and the response returned, which is what makes a trace useful for diagnosing an incorrect answer. Access to traces is limited to Edmund engineering personnel, and no trace is held beyond the retention period stated above.
Telemetry and observability
Edmund operates its own observability stack rather than subscribing to a hosted monitoring service. The components below are open-source software deployed on infrastructure Edmund manages. Grafana Labs, which publishes that software, receives no data and is not a sub-processor. The anonymous usage statistics these components would otherwise report about themselves — versions in use, feature counts — are disabled in the configuration of every component.
| Component | Purpose | Hosting | Retention |
|---|---|---|---|
| Grafana Alloy | Telemetry collection agent | Edmund-managed Kubernetes on AWS, Ireland, eu-west-1 (production); Germany, eu-central-1 (staging) | None — forwards only |
| Loki | Application and platform logs | Edmund-managed Kubernetes on Azure, Netherlands, westeurope | 14 days |
| Mimir | Infrastructure and application metrics | Edmund-managed Kubernetes on Azure, Netherlands, westeurope | 14 days |
| Tempo | Distributed traces | Edmund-managed Kubernetes on Azure, Netherlands, westeurope | 14 days |
| Grafana | Dashboards and alerting | Edmund-managed Kubernetes on Azure, Netherlands, westeurope | None — queries the components above |
The storage components persist to Microsoft Azure Blob Storage in the Netherlands (westeurope), so all telemetry remains within the European Union. Azure appears as a sub-processor above for that reason, and is covered by the Microsoft Products and Services Data Protection Addendum.
Deletion is enforced by configuration rather than by a manual process, in two layers. The components themselves delete telemetry once it reaches 14 days old; the storage account then hard-deletes any object that survived that pass within 16 days, and its soft-delete window keeps a deleted object recoverable for a further 7 days. The 14 days above is therefore the retention Edmund operates to, not the last moment at which a copy can physically exist.
Loki receives the logs of every pod in the cluster unless that pod opts out, together with Kubernetes events. Log lines are neither filtered nor redacted, so personal data that an application writes into its own logs is retained for the period stated above, and access to it is restricted to Edmund personnel.
Infrastructure metrics — those scraped from the nodes, containers and cluster services — are gathered against an allowlist of metric names, so a newly exposed metric is not collected until it is added deliberately. Those metrics are counters, gauges and timings, and carry no identifiers. Metrics that the platform’s own applications emit are forwarded without that allowlist.
Access to Grafana is restricted to Edmund personnel and authenticated through Microsoft Entra ID single sign-on. Dashboards are further restricted by role.
For On-premise deployments, telemetry collection is optional and configured as part of the installation; where it is disabled, none of the above leaves the customer’s own infrastructure.
Internal analytical access
Edmund operates a self-hosted Redash instance for internal analysis of platform data — usage patterns, data quality, and the questions that arise when supporting a deployment. Like the observability stack it is open-source software running on infrastructure Edmund manages, not a hosted service, and no third party is involved.
| Property | Configuration |
|---|---|
| Hosting | Edmund-managed Kubernetes on Azure, Netherlands, westeurope |
| Authentication | SAML single sign-on against Microsoft Entra ID; password login disabled |
| Reachable by | Edmund personnel only |
| Data reached | Edmund’s production and staging databases, over a private WireGuard tunnel |
| Result storage | Redash’s own PostgreSQL database on Azure, Netherlands, westeurope |
| Backup retention | 7 days, without geographic replication |
Two properties of this arrangement matter to a customer reading this policy.
The databases Redash queries are not exposed to the public internet, and Redash does not reach them across one. Access runs over a private WireGuard tunnel to the network the databases sit in, and Redash’s own database is likewise closed to public network access and reachable only from within Edmund’s virtual network.
A query result, however, is stored by Redash so that a dashboard can be re-opened without re-running the query. Where a query returns customer data, that result is a copy of customer data held outside the database it came from, for as long as the cached result exists. It is subject to the same access restriction as the rest of the instance, and it is removed by Redash’s periodic clean-up of results no longer referenced by a saved query.
For On-premise deployments Redash has no route to the customer’s infrastructure, and no platform data leaves it by this path.
Policy review
This policy is reviewed annually, and additionally whenever legal requirements, business operations, or the platform’s technology change.