Enterprise Distributed Memory Requirements
An enterprise's agents run in regions, departments, data centres and on laptops and need a shared memory, but one central store is a single point of control, failure and exposure, and some sites must keep some data at home. Everything Memory runs one engine per site and peers the engines into a mesh with no central store; sends a peer only what a site shares and the peer accepts, policy by policy; passes on what arrives under its own shares; and keeps every engine working when it is cut off.
Approach
- One engine per site. Each site runs its own engine, with its own database and its own archive. Agents at the site talk to it.
- Engines peer into a mesh. Engines connected to one another form a mesh. There is no central store and no hub.
- What crosses is chosen per peer. A site shares a policy, or all of them, with a peer; the peer accepts a policy by its id. Only what is both shared and accepted crosses: the policy first, then its lanes, whole, with their records. Nothing crosses by default.
- What arrives is passed on. An engine passes on what it received to its own peers, under its own shares.
- Each engine works on its own. An engine reads and writes with no connection to its peers, and exchanges what it missed when it reconnects. An air-gapped site runs fully offline, and exchanges only if and when it is connected.
Example
The EMEA engine holds two policies: sales, with the lane sales_emea, and legal, with the lane legal_emea. It shares sales with headquarters, and headquarters accepts it. Records in sales_emea cross to headquarters, and headquarters' agents read them like their own, within their agent profiles. legal is not shared, so legal_emea and its records never leave the region.
Reference
The mesh, peering and exchange: mesh.