Clause 14.2
14.2 Data location. The Provider shall store and process Customer Documents, including signed documents, their audit trails and all backups of them, only within the European Union.
Kessler Maschinenbau, a machine-tool maker near Stuttgart, would not sign the e-signature service's enterprise agreement without that clause, and sales agreed to it. The service runs in one region, in Virginia. Its customers are in the EU, the US and Australia, and until this week every signed contract from every one of them sat in the same Postgres cluster and the same storage bucket.
The clause reads like one requirement and is three questions: where each request goes, where each customer's data lives, and what happens to both when a region dies. Teams get into trouble by answering them as one. A design that answers only the first ends up fast and in breach of contract; one that answers only the second ends up compliant and slow.
Nearest is about speed
Distance alone justifies more than one region. A round trip across an ocean costs roughly 70–150 ms before your code runs at all, and Sydney to Virginia is farther still, so an Australian signer whose page needs a handful of sequential round trips waits a second for nothing. The service adds regions in Frankfurt and Sydney, and GeoDNS or anycast sends each user to the nearest healthy one. DNS, proxies & the network edge explains how those work, and why a DNS answer is only a guess about where the user is. The signing page's scripts, fonts and images come from a CDN even closer (CDN & static delivery).
Proximity says nothing about where Kessler's documents may be. Kessler's purchasing manager, signing a supplier contract from a trade fair in Chicago, is nearest to Virginia. If the nearest region simply served her, her signed contract would be written to a US database. Routing decides where a request enters; something else has to decide where it is served.
A home region for every account
That something is a home region: each customer account gets one, chosen by proximity or by law, and its documents, signatures, audit trails and backups never leave that region's jurisdiction. Kessler's home is Frankfurt, in the EU. A constraint like clause 14.2 is data residency: a contract or a law that says in which jurisdiction data must stay. Splitting data by account and region is partitioning at the scale of a whole region, with the account as the key (Sharding & partitioning covers choosing keys). Inside a region, accounts still share tables, and keeping them apart is the business of Multi-tenancy.
A request finds its home through a small, global account directory that maps each account to its home region. Sign-in reads it once and puts the home region in the session token, so every later request carries it. The Chicago request enters Virginia, the US app reads EU in the token and forwards the whole request to Frankfurt. The US app stores nothing, but it decrypts the request on the way. If the contract counts that as processing, the switch has to happen before any document is sent: sign-in points the browser at Frankfurt's own address, and the requests that carry documents go there directly. Signing links sent by email carry the region in the envelope id, so a signer who never logged in reaches the right region without a directory lookup.
GeoDNS sends each user to the nearest region. The US app forwards requests for EU-homed accounts to the EU app; Kessler's documents exist only in the EU database, and no edge copies them out.
Forward the whole request, not each query. If the US app rendered the page itself and fetched Kessler's data from Frankfurt one query at a time, five sequential cross-region reads at about 100 ms a round trip would add half a second on the wire alone. A cheaper-looking design hides the same trap: app servers in every region all writing to one primary in Virginia. The ones outside the US still cross the ocean on every write, and on every read that no local replica or cache can answer. The deployment shapes have patterns of their own: Geode, where any region serves any user, a promise residency narrows, and Deployment Stamps, complete stacks for groups of customers that can sit in the region their data must stay in.
What may cross the ocean
The directory is global, so it holds nothing clause 14.2 covers: account ids, the home region, the plan. Whether a global sign-in index may hold EU users' email addresses is a question for the privacy team. Everything Kessler owns stays in the EU, and the leaks found in practice are rarely the main database:
- Backups and copies. A nightly backup copied to a Virginia bucket for safety is a copy of every signed document, outside the EU. Kessler's backups go to a second EU location.
- Side paths. Logs that include document titles shipped to a US logging service, one search index for all customers, a support tool in Virginia that queries Frankfurt and caches what it read: each moves Kessler's data without anyone deciding to.
- Global jobs. Billing needs how many envelopes Kessler sent, not what was in them, so a job in Frankfurt sends counts to the global billing service, and no job in Virginia reads Frankfurt's rows.
- Reads of another region's data. A request that really needs data homed elsewhere either pays the latency openly, as a labelled cross-region call, or reads a local replica and accepts that it may be stale (Replication & read replicas; for what a reader may see, Consistency models: strong vs eventual (CAP)). For residency-bound data a replica is not an option: a copy in Virginia is the breach.
Some crossings are not engineering questions. When a supplier in Ohio opens Kessler's signing link, the document travels to the supplier's browser; residency clauses are usually read as covering where the provider stores and processes data, not whom the customer sends it to, but that reading belongs to the lawyers.
Kessler's subsidiary in Melbourne signs through the same account, so its staff enter in Sydney and are forwarded to Frankfurt on every request: their documents are Kessler's documents. Only if those contracts may live in Australia can the subsidiary become a second account with its own home, and then whatever the two share, a template library or a contract both sign, has to be copied deliberately or reached across regions.
The night Frankfurt goes dark
When a region fails, its users are rerouted, and how fast depends on where their data already is. Either it is already replicated to another region (active-active or a warm standby: fast, expensive) or it is restored from backups (slow, cheap), and residency-bound data must fail over inside its jurisdiction. Virginia is no failover target for Kessler, however healthy it is: serving Kessler from there means Kessler's documents are there. So the EU home is really a pair, Frankfurt plus a second EU region with a warm standby, while US and Australian accounts without residency terms can fail over wherever there is capacity.
Promoting the standby without ending up with two primaries is the job of Failover, and asynchronous replication can drop the last writes when it happens (Replication & read replicas has the details). If region failover has never been executed, you don't have failover; you have a diagram of one, and Chaos engineering & game days is about rehearsing it.
What still crosses, and who decides
Home regions are cheap to choose at sign-up and expensive to add later: moving EU customers' data out of a global table after the fact can be a migration project measured in quarters, with the copy, the cutover and every side path above to find. The Multi-Region Active-Active Challenge asks you to explain which data may be regional-only for residency reasons, next to accepting writes in every region and choosing a conflict rule. Answer it on the Grid: put the region in the title of every residency-bound database, and label each edge that leaves a region with what it carries.
What crosses is now a short list, each item on a labelled edge: the directory, billing counts, a signer's browser, a support engineer's screen. Whether each one may cross is for the contract and the lawyers on both sides to settle, and so is the hardest case, a jurisdiction served by a single region: when that region fails, the choice is to stay down or to break the promise, and the contract should say which before the outage, not a runbook during it.