METAFILES by V-Connect

Docs / Platform

Architecture

The servers of a METAFILES site, what runs on each, and how files, document edits and calls travel through it.

A METAFILES site is made of a small number of servers in three tiers, an encrypted edge in front of them and a control plane beside them. This page describes each part and then follows a request through the site.

The servers and their roles

Role How many What runs there
App servers 2 The load balancer, one METAFILES portal per organisation, the storage engine (active on one server, standby on the other), a multi-master database cluster and a cache with automatic promotion.
Office servers 2 Collabora Online for document editing, the Talk services for calls (signalling, a media server and a private call relay), and the site's monitoring.
Storage servers 3 The Ceph storage cluster, which keeps three copies of every object, with its S3 gateways; the site's secret vault, which unseals itself automatically; and the internal certificate authority that issues every service certificate.
Public call relays 2 Relays that carry call audio and video for people outside the site's network.
Control plane separate Builds organisations, carries out approved machine actions, observes the site every 10 minutes and checks it against its intended state every day.

Each tier has a floating site address that moves to a healthy server automatically: one for the web, one for storage and one for calls.

The edge

All web traffic reaches the site through an encrypted tunnel, so the site opens no inbound web ports to the internet. One tunnel connector runs on each app server, and either can carry the traffic alone. Only known names and routes pass the edge; anything else is refused.

The edge also carries per-organisation rules written by the console: a suspended organisation gets the "This service is suspended" page, and a read-only organisation has every change refused before it reaches the portal.

The portals

Every organisation has its own METAFILES portal, running on both app servers at the same time. The load balancer spreads users across both. Each portal has its own database, its own cache space and its own route, and its configuration is identical on both servers.

Services on each app server keep every portal in line with the settings in the MSP console (branding, sharing, two-factor, quotas, apps, users) and re-apply them every few minutes.

The engine

DMSVault's storage engine is the core of the product. It is written in Rust and does the real work with every file:

  • splits files into chunks and stores identical chunks only once (content-addressed deduplication);
  • compresses and encrypts every chunk with the organisation's own key;
  • keeps versions, trash and restore points, and cleans up what is no longer needed (garbage collection);
  • enforces quotas, read-only mode, WORM, legal hold and ransomware protection at its S3 gateway, which every portal byte passes through;
  • serves the MSP console and keeps each organisation's tamper-evident audit trail.

The portals talk to the engine only through its S3 gateway; they never see the storage cluster. Nothing writes to the store directly except the engine.

The engine runs on one app server with a standby on the other. Safeguards make sure that only one engine can ever write: the standby has no access to storage until it is promoted, and a promoted engine first locks out the old one. See Scaling and high availability.

Storage

On the site's own storage, the engine writes to a Ceph cluster on the three storage servers. Every object is stored three times, once per storage server, on encrypted disks, and the cluster keeps serving with two copies while one server is down. The engine reaches the storage gateways over mutually authenticated TLS and balances across all three.

An organisation can also keep its encrypted data elsewhere: in Amazon S3, in another S3-compatible service or on its own server. See Storage.

How requests travel

Saving a file

  1. The browser or METAFILES client sends the file to https://<organisation>.dmsvault.cloud. Large files are sent in chunks.
  2. The edge checks the organisation's access rules and passes the request to its portal on either app server.
  3. The portal stores the file through the engine's S3 gateway.
  4. The engine chunks, deduplicates, compresses and encrypts it with the organisation's key, checking quota, WORM, legal hold, read-only and ransomware protection.
  5. The encrypted data is written to the organisation's store: on the site's storage, three copies on three servers.
  6. The engine confirms only after the write is safely recorded, and the user sees the file.

Reading a file

The portal asks the engine; the engine reads the encrypted data from any of the storage gateways, decrypts and reassembles it, and streams it back. Reading works even while an organisation is read-only.

Editing a document

The portal opens Collabora Online in the browser. Everyone editing the same document is routed to the same office server, so they work on one copy together. Collabora reads and saves the document through the portal, and each save reaches the engine as a new version. A read-only organisation cannot save.

A call

Talk hands each participant a signalling address and short-lived relay credentials. Signalling runs on the office servers, and a media server forwards streams in group calls. People on the site's own network use the private relay; everyone else uses the two public relays, which refuse to relay towards private addresses.

Building an organisation

An MSP's Add organisation is recorded by the engine as a request. The control plane picks it up and runs the 12 build steps one organisation at a time, each previewed, applied and then re-run to confirm nothing more changes, and ends with an isolation check. The console shows every step live. See Organisations.

The control plane

The control plane is separate from the site's servers. It holds the signed release of the deployment kit and checks it before and after every run. It builds and deletes organisations, carries out approved machine actions, and observes the site.

Users never depend on it while they work. If the control plane is unavailable, files, sign-ins, Office and calls carry on; only new builds, deletes and machine actions wait until it is back.