Three pulls away from one region
A payments company in Lagos started the way the earlier lessons recommend: one provider, one region — Frankfurt, the nearest with every service it needed — and three availability zones. Five years on it serves customers in Nigeria, Kenya, India and Germany, and three complaints arrive in the same quarter.
Customers in Lagos say the app feels sluggish, and the measurements agree: every page waits for several round trips to Germany. The board asks what happens if the whole provider goes down, as large parts of several providers did at least once in 2025. And the compliance team reports that India's central bank wants payment data stored only in India, while European customers' data is governed by European rules on transfers abroad.
Each complaint pulls the system out of its single region in a different direction: towards the edge, closer to users; towards more than one cloud; and towards sovereignty, keeping data and control within particular borders. This lesson takes them in turn, with the numbers that tell you when each pull is worth following. It builds on cloud networking for latency, on storage and databases for replication, and on what the cloud is for the arithmetic of correlated failures.
The edge, and CDNs that run code
Definition— Edge computing
Running computation and storage in many small sites close to users or devices — a content-delivery network's points of presence, a telecom operator's facilities, a factory or a shop — instead of, or in front of, a few large regions. The origin is the central system the edge stands in front of.
A content-delivery network, as cloud networking showed, caches content at hundreds of sites, terminates TCP and TLS near the user, carries misses to the origin over its own backbone on connections it keeps open, and absorbs floods of traffic.
Since the late 2010s, CDNs have also run customer code. A small function runs at the edge site on each request: rewriting URLs, checking a token, choosing a language, assembling a page from cached parts, or answering an API call outright. The platforms differ in what they isolate and allow, each described in its own documentation: Cloudflare Workers run JavaScript and WebAssembly in V8 isolates; Fastly Compute runs WebAssembly; Akamai EdgeWorkers run JavaScript; Amazon CloudFront offers lightweight CloudFront Functions and the heavier Lambda@Edge; Google Cloud's Service Extensions run WebAssembly plugins in its load balancers and CDN; Alibaba Cloud has Edge Routine. The isolation technologies underneath — isolates and WebAssembly — are the subject of the next lesson.
Edge functions come with limits that follow from running everywhere at once: tight caps on CPU time and memory per request, restricted APIs, short or no background work, and above all distance from your data. Edge platforms offer storage — key–value stores, object stores, small SQL databases — but replicated to hundreds of sites it is usually eventually consistent, or else every write goes to one primary location.
Further out is the far edge: compute in telecom networks (multi-access edge computing on 5G), provider hardware installed in a customer's own building (AWS Outposts, Azure Local, Google Distributed Cloud), and fleets of small Kubernetes clusters in shops, factories and vehicles managed with projects such as K3s and KubeEdge. These exist for latency, for bandwidth (a factory's cameras produce more video than it is worth sending away) and for the ability to keep working when the link to the region fails.
Latency maths for edge placement
The fundamentals are in cloud networking's latency budgets: light in fibre covers about 200 km per millisecond, cable routes run 1.3 to 2 times the great-circle distance (computed with the haversine formula of spherical geometry, non-Euclidean geometry), and a new connection spends three round trips — TCP, TLS 1.3 (RFC 8446) and the request — before the first byte arrives. QUIC (RFC 9000) merges the first two. Here we apply them to the two questions an edge design must answer: when does code at the edge pay, and where should the sites go?
The block takes a user in Lagos and a database in Frankfurt, and compares running the request's code at the origin, next to the database, with running it at an edge site in Lagos, as the number of sequential database queries per request grows — first on a new connection, then on one the user's browser already holds open.
import math
def km(p, q):
"""Great-circle (haversine) distance on a sphere of radius 6371 km."""
(a, b), (c, d) = [(math.radians(x), math.radians(y)) for x, y in (p, q)]
h = math.sin((c - a) / 2) ** 2 + math.cos(a) * math.cos(c) * math.sin((d - b) / 2) ** 2
return 2 * 6371 * math.asin(math.sqrt(h))
LAGOS, FRANKFURT = (6.52, 3.38), (50.11, 8.68)
far = 2 * km(LAGOS, FRANKFURT) * 1.5 / 200 + 1.0 # round trip Lagos-Frankfurt: detour 1.5, 200 km/ms, +1 ms
near = 4.0 # round trip to an edge site in the user's own city
in_zone = 0.5 # a database query from a server in the database's zone
server = 20.0 # computation per request, ms
def response_ms(code_at, queries, new_connection):
"""One request that makes `queries` sequential database calls; the database is in Frankfurt."""
trips = 3 if new_connection else 1 # TCP + TLS 1.3 + request, or just the request
if code_at == "origin":
return trips * far + queries * in_zone + server
return trips * near + queries * far + server # edge: handshakes are local, every query crosses
print(f"round trip Lagos-Frankfurt: {far:.0f} ms\n")
print(f"{'queries':>7s} {'new connection':>22s} {'reused connection':>22s}")
print(f"{'':7s} {'origin':>10s} {'edge':>11s} {'origin':>10s} {'edge':>11s}")
for q in range(5):
row = [response_ms(w, q, new) for new in (True, False) for w in ("origin", "edge")]
print(f"{q:7d} {row[0]:7.0f} ms {row[1]:8.0f} ms {row[2]:7.0f} ms {row[3]:8.0f} ms")
# An edge cache in front of origin code: hits are answered locally, misses go to the origin.
for hit_ratio in (0.5, 0.9, 0.99):
hit = near + 1.0 # reused connection, answer from cache
miss = near + far + 3 * in_zone + server # one trip to the origin, which queries nearby
print(f"edge cache with {hit_ratio:.0%} hits, reused connection: {hit_ratio * hit + (1 - hit_ratio) * miss:5.1f} ms on average")On a new connection, edge code wins while it makes at most two queries to Frankfurt — the local handshakes save two long round trips, and each query spends one. On a connection that is already open there are no handshakes to save, and a single query to the origin makes edge code slower than origin code (98 ms against 95). Caching changes the picture completely: an edge cache answering 90% of requests brings the average to about 15 ms. Edge code pays when it needs no data from the origin, or when the data it needs is replicated to the edge.
Where to put the sites
An edge network is a facility-location problem: choose sites from a set of candidates to minimise the average latency of users, each served by the nearest site, weighted by how many users there are. Finding the best exactly is NP-hard in general (the -median problem; see optimisation), but small cases can be solved by trying every set, and a greedy rule — add whichever site helps most, one at a time — is a common heuristic. The block compares the two for twelve African cities with synthetic user counts.
import math
from itertools import combinations
# Where should a service with users across Africa put its edge sites? Approximate coordinates;
# the user weights are synthetic (thousands of daily users), not real statistics.
USERS = {
"Lagos": ((6.52, 3.38), 900), "Cairo": ((30.04, 31.24), 700), "Kinshasa": ((-4.32, 15.31), 400),
"Johannesburg": ((-26.20, 28.05), 600), "Nairobi": ((-1.29, 36.82), 500), "Addis Ababa": ((9.03, 38.74), 350),
"Accra": ((5.60, -0.19), 300), "Casablanca": ((33.57, -7.59), 350), "Dakar": ((14.72, -17.47), 200),
"Dar es Salaam": ((-6.79, 39.21), 300), "Luanda": ((-8.84, 13.23), 250), "Khartoum": ((15.50, 32.56), 150),
}
SITES = ["Lagos", "Cairo", "Johannesburg", "Nairobi", "Casablanca", "Accra", "Kinshasa", "Dakar"]
def km(p, q):
(a, b), (c, d) = [(math.radians(x), math.radians(y)) for x, y in (p, q)]
h = math.sin((c - a) / 2) ** 2 + math.cos(a) * math.cos(c) * math.sin((d - b) / 2) ** 2
return 2 * 6371 * math.asin(math.sqrt(h))
def rtt(user, site):
return 4.0 if user == site else 2 * km(USERS[user][0], USERS[site][0]) * 1.5 / 200 + 1.0
def mean_rtt(sites):
"""Each user goes to its nearest site; average over users, weighted by how many there are."""
total = sum(w for _, w in USERS.values())
return sum(w * min(rtt(u, s) for s in sites) for u, (_, w) in USERS.items()) / total
greedy = []
print(f"{'k':>2s} {'greedy: add the best site each time':46s} {'RTT':>7s} {'best set of k sites (all tried)':46s} {'RTT':>7s}")
for k in range(1, 6):
greedy.append(min((s for s in SITES if s not in greedy), key=lambda s: mean_rtt(greedy + [s])))
best = min(combinations(SITES, k), key=lambda c: mean_rtt(list(c)))
print(f"{k:2d} {', '.join(greedy):46s} {mean_rtt(greedy):4.1f} ms {', '.join(best):46s} {mean_rtt(list(best)):4.1f} ms")
frankfurt = (50.11, 8.68)
total = sum(w for _, w in USERS.values())
print(f"\neverything served from Frankfurt: {sum(w * (2 * km(p, frankfurt) * 1.5 / 200 + 1) for p, w in USERS.values()) / total:.0f} ms")One site in Kinshasa, near the middle of the continent, beats any other single site, and it cuts the average round trip from 81 ms (everything from Frankfurt) to 40 ms. But the best pair is Lagos and Nairobi, without Kinshasa: greedy's first choice was a compromise that later sites make redundant, and greedy stays about 3 ms behind the optimum at and . Real planners start from greedy and then improve by swapping sites in and out, which closes most of the gap. The returns also diminish: the fifth site saves less than 5 ms on average.
Multi-cloud and hybrid, and what they really cost
Definition— Multi-cloud and hybrid cloud
A multi-cloud system uses public clouds from more than one provider. It may put different workloads on different providers (the most common form), keep a standby copy on a second provider, or run active-active across providers so that either can carry all traffic. A hybrid system combines a public cloud with the organisation's own data centres or private cloud, connected so that workloads and data can move between them.
Organisations go multi-cloud for several reasons, some more convincing than others:
- Inheritance. Most multi-cloud estates were not designed: a merger brought a second provider, or one team chose differently.
- Best fit. A particular database, AI model, accelerator or regional presence exists on only one provider.
- Regulation and sovereignty. A required region or certification exists only on some providers — the subject of the next section.
- Negotiating power and exit. A credible ability to leave keeps prices honest. Financial regulators increasingly ask for exit plans; the EU's Digital Operational Resilience Act (DORA), applying since January 2025, requires them of financial firms for critical outsourced services.
- Resilience against a whole provider failing. This is the reason given most often, and the one that most needs numbers.
Whole-provider failures are rare, but they happen. In 2025 a latent race condition in the DNS management of one regional database service at AWS caused a long outage in its largest region, felt worldwide through services that depended on it (AWS summary); a policy update that triggered a latent bug in the system that checks API requests disrupted many Google Cloud products across regions for hours (Google Cloud report); and an automatically generated configuration file that had grown too large crashed much of Cloudflare's network (Cloudflare report). Each published a detailed account, which is good practice worth reading. Notice the common thread: a global control-plane or configuration dependency, exactly the common-mode failure that zones and regions cannot escape (what the cloud is).
The block puts illustrative numbers on three designs: the payments company's current single region; two regions of one provider, with failover; and two providers, active-active. Each design suffers the incidents we cause ourselves; each escapes some of the provider's.
# Expected downtime and cost per year for three designs. Every number is an illustrative
# assumption: replace them with your own incident history and your own bills.
HOURS = 8760
region_outages, region_hours = 1.0, 3.0 # per year: a region-wide incident, and how long it lasts
provider_outages, provider_hours = 0.3, 4.0 # an incident in a provider's global control plane or identity service
own_incidents, own_hours = 4.0, 1.0 # incidents we cause ourselves: bad deploys, expired certificates...
base_infra = 600_000 # USD per year for one region of the service
changes_gb_per_day = 2_000 # data that must be replicated to the other site each day
designs = {
# name: (failover minutes for a region outage, for a provider-wide one (None = no escape),
# infrastructure multiplier, USD per GB replicated, extra people-time per year in USD)
"one region, three zones": (None, None, 1.0, 0.00, 0),
"two regions, one provider": (15, None, 1.6, 0.02, 100_000),
"two providers, active-active": (2, 2, 2.1, 0.08, 300_000),
}
previous = None
print(f"{'design':30s} {'downtime h/yr':>13s} {'availability':>12s} {'cost USD/yr':>12s} {'USD per hour saved':>19s}")
for name, (fo_region, fo_provider, mult, egress, people) in designs.items():
down = own_incidents * own_hours # our own mistakes hit every design
down += region_outages * (region_hours if fo_region is None else fo_region / 60)
down += provider_outages * (provider_hours if fo_provider is None else fo_provider / 60)
cost = base_infra * mult + changes_gb_per_day * 365 * egress + people
# what each hour of downtime avoided costs, compared with the design above it
ratio = "-" if previous is None else f"{(cost - previous[1]) / (previous[0] - down):,.0f}"
previous = (down, cost)
print(f"{name:30s} {down:13.2f} {1 - down / HOURS:12.3%} {cost:12,.0f} {ratio:>19s}")Under these assumptions a second region removes most of the region-level downtime for about USD 173,000 per hour saved; a second provider then removes the provider-wide term, 1.2 hours a year, at about USD 387,000 per hour saved. Whether that is worth it depends entirely on what an hour of downtime costs the business — for a payments company in a busy season, it might be. Notice too what remains: four hours a year of the company's own incidents, now the largest term in every design.
The real costs
- Egress. Providers charge for data leaving their network. Example list prices for internet egress from the largest providers were around USD 0.05–0.12 per GB in 2025–26, depending on volume and destination, against about USD 0.01–0.02 between their own regions on one continent; the second provider in the block costs four times as much in replication alone. In 2024 the three largest providers began waiving egress fees for customers moving their data out entirely, and the EU Data Act requires providers to phase out switching charges by 12 January 2027 — but egress for running in parallel on two providers may still be charged, at cost.
- The lowest common denominator. To run on two providers, a system can use only what both offer, or must maintain two implementations. Managed databases, queues, identity systems and AI services differ in semantics, not just names.
- People. Two sets of skills, tools, security reviews, on-call runbooks and contracts. This is usually the largest cost, and the hardest to see on a bill.
- Data gravity. Large datasets are slow and expensive to move, so compute tends to follow them; "portable" applications pinned to immovable data are not portable.
Portability layers help with some of this and not with the rest. Kubernetes gives the same deployment interface on every provider (Kubernetes), and one infrastructure-as-code tool can manage several providers (infrastructure as code) — but the resources remain provider-specific, as two storage buckets for the same backups show (OpenTofu or Terraform, not run here):
# Not run here: OpenTofu / Terraform. One tool, two providers, two different resource types.
provider "aws" {
region = "eu-central-1"
}
provider "google" {
project = "lagos-payments-dr"
region = "europe-west3"
}
resource "aws_s3_bucket" "backups" {
bucket = "lagos-payments-backups-primary"
}
resource "google_storage_bucket" "backups" {
name = "lagos-payments-backups-secondary"
location = "EUROPE-WEST3"
uniform_bucket_level_access = true
}The pragmatic pattern in 2026 is selective: one primary provider used deeply; a second for specific workloads, regions or exit readiness; backups copied to a second provider or location the primary account cannot delete; and active-active across providers only for the few services whose downtime is ruinously expensive.
Data residency and digital sovereignty
Several terms are often used interchangeably and should not be.
Definition— Residency, localisation and sovereignty
Data residency is where data is physically stored and processed. Data localisation is a legal requirement that some data stay within a jurisdiction, or that a copy be kept there. Data sovereignty is the principle that data is subject to the laws of the jurisdictions that can reach it — where it is stored, but also where the company holding it, or its parent, is based. Digital sovereignty, more broadly, is a state's or organisation's ability to control its own data, infrastructure and technology choices, including keeping services running if a foreign supplier withdraws.
The difference matters because the laws differ. Some restrict where data may go; some require it to stay; and some give governments access to data held by their companies wherever it is stored. The table summarises the main rules in several jurisdictions as of November 2026. It is a map for engineers, not legal advice; these rules change, and their details depend on the sector and the kind of data.
| Jurisdiction | Main instruments | What they mean for where data may go |
|---|---|---|
| European Union and EEA | GDPR (applies since May 2018); Data Act (applies since September 2025) | No general rule that personal data must stay in Europe, but a transfer outside the EEA — including remote access from abroad — needs a basis: an adequacy decision for the destination, or safeguards such as standard contractual clauses with an assessment of the destination's laws. US recipients certified under the EU–US Data Privacy Framework (adequacy decision of July 2023) qualify; the EU General Court upheld the framework in September 2025, and an appeal to the Court of Justice is pending. The Data Act adds switching rights and obliges providers to resist unlawful foreign government access to non-personal data. |
| France | ANSSI's SecNumCloud qualification | A demanding security qualification whose current version also requires protection from non-European laws; French public-sector policy requires it for some sensitive workloads. |
| Germany | BSI's C5 criteria catalogue | An audited attestation of cloud security controls, including transparency about data location and jurisdiction, widely required by public bodies and regulated firms. |
| India | Digital Personal Data Protection Act 2023, with Rules notified in November 2025 and phased in over 18 months; Reserve Bank of India directive of 2018 | Transfers abroad are allowed except to countries the government restricts. Sector regulators can be stricter: the central bank requires payment-system data to be stored only in India. |
| China | Cybersecurity Law (2017), Data Security Law (2021), Personal Information Protection Law (2021) | Operators of critical information infrastructure, and handlers of large volumes of personal information, must store it in China; exports need a security assessment, a standard contract or certification, with some thresholds relaxed in 2024. Foreign providers operate in China through licensed local partners. |
| Russia | Federal Law 242-FZ (in force since 2015) | Personal data of Russian citizens must be recorded and stored, at least initially, in databases in Russia. |
| Indonesia | Government Regulation 71 of 2019 | Private-sector systems may be hosted abroad; public-sector systems and data generally must stay in Indonesia. |
| Kenya | Data Protection Act 2019 and its 2021 regulations | Processing for listed purposes of strategic state interest must take place in Kenya, or keep a serving copy there; other transfers need safeguards. |
| Brazil | LGPD (in force since 2020); the data-protection authority's 2024 regulation on international transfers | No general localisation; transfers need a basis such as adequacy or approved standard clauses. |
| United States | No general localisation law; the CLOUD Act (2018); a Justice Department rule on bulk sensitive data (in force since April 2025); FedRAMP for federal agencies | US law can require providers subject to it to disclose data in their possession, custody or control wherever it is stored — the main reason other jurisdictions talk about sovereignty. The Justice Department rule restricts some transfers of bulk sensitive personal data to designated countries of concern. |
For the payments company this means: Indian payment data must live in Indian regions, including backups; European customer data may stay in Frankfurt and may leave only with a basis; and every copy counts, including the ones nobody thinks of as "the database". The block audits an inventory of copies against the company's own policy.
# Does every copy of the data sit where policy allows? A fictional company's inventory of data copies,
# checked against its own residency policy (drawn up with its lawyers; simplified, and not legal advice).
EEA = {"eu-frankfurt", "eu-paris", "eu-dublin", "eu-stockholm"}
POLICY = { # data class -> (regions allowed outright, safeguards that make another region acceptable)
"card-payments-india": ({"in-mumbai", "in-hyderabad"}, set()), # sector rule: store only in India
"customers-eu": (EEA, {"adequacy decision", "standard contractual clauses"}),
"telemetry-anonymous": (None, set()), # truly anonymous: no restriction
}
COPIES = [ # (what, data class, region where it is stored or accessed from, safeguard on file)
("payments primary database", "card-payments-india", "in-mumbai", None),
("payments failover replica", "card-payments-india", "in-hyderabad", None),
("payments nightly backup", "card-payments-india", "ap-singapore", "standard contractual clauses"),
("customer database", "customers-eu", "eu-frankfurt", None),
("customer read replica", "customers-eu", "eu-paris", None),
("application logs (contain emails)", "customers-eu", "us-virginia", "adequacy decision"),
("support desk remote access", "customers-eu", "ph-manila", None),
("product analytics", "telemetry-anonymous", "us-oregon", None),
]
def check(cls, region, safeguard):
allowed, safeguards = POLICY[cls]
if allowed is None or region in allowed:
return "ok"
if safeguard in safeguards:
return "ok with " + safeguard
return "VIOLATION"
for what, cls, region, safeguard in COPIES:
print(f"{check(cls, region, safeguard):32s} {what:36s} {region}")Two violations. The payment backup in Singapore breaks a rule that no contract can satisfy; it must move to an Indian region. The support desk in Manila reading European customer records is a transfer too, and needs a basis that has not been documented. The logs in Virginia are acceptable only because the logging vendor is certified under the Data Privacy Framework — a basis that would vanish if the courts struck the framework down, as they did its two predecessors in 2015 and 2020.
Sovereign cloud offerings
Providers have answered with offerings that differ mainly in who operates the cloud and which law the operator answers to. As of late 2026, side by side:
- Sovereign variants of global clouds. The AWS European Sovereign Cloud, physically and logically separate infrastructure operated by EU-resident staff, opened its first region, in Brandenburg, Germany, in January 2026. Microsoft completed its EU Data Boundary in 2025 and announced sovereign offerings that June: sovereign controls in its public cloud, a private option on Azure Local in customer or partner data centres, and national partner clouds operated by local companies, such as Bleu in France and Delos Cloud in Germany. Google Cloud offers data-boundary controls, partner-operated clouds (S3NS, controlled by Thales, in France; with T-Systems in Germany) and an air-gapped Google Distributed Cloud. Oracle has an EU Sovereign Cloud and dedicated regions.
- Locally owned providers, which are subject only to their home jurisdiction: in Europe, for example, OVHcloud, Scaleway, IONOS, STACKIT and Open Telekom Cloud; in China, domestic providers such as Alibaba Cloud, Huawei Cloud and Tencent Cloud; elsewhere, national providers and government clouds. Their service catalogues are generally narrower than those of the largest global providers.
- Your own infrastructure, private clouds built on open-source stacks such as OpenStack and Kubernetes, which give the most control and the most work.
European governments and companies have also built shared frameworks, such as Gaia-X for trusted data sharing, which is a set of rules and labels rather than a cloud.
Whatever the label, the same questions decide the trust model. Which company operates the service, and under which jurisdiction is it and its parent? Who holds the encryption keys — can you keep them in your own hardware security modules? Which staff can access systems, from where, and with whose approval? Can the service keep running if disconnected from the parent company? How far behind the global service are its features, and at what price premium? An honest answer to "is it sovereign?" is usually "sovereign against some threats and not others".
The professional view
- Cache first, compute at the edge second. Most of the latency gain comes from caching and terminating connections near users; move code to the edge only with the data it needs, or with requests that need no data.
- Measure from users. Real-user monitoring of round trips beats any map; check placements against it.
- Escape a region before escaping a provider. Multi-region within one provider removes most outage risk at lower cost; go multi-provider for specific reasons — a regulator, an exit plan, a workload — and keep immutable backups somewhere the primary account cannot delete.
- Price egress and people before committing to active-active across providers, and rehearse failover, or it will not work when needed.
- Keep a data inventory that lists every copy — replicas, backups, logs, analytics, caches, support access — with its location and legal basis, and check it automatically on every infrastructure change.
- Involve lawyers early and date your assumptions. The rules above changed repeatedly between 2015 and 2026; design so that data can move when they change again.
Common misconceptions
Misconception
"Running code at the edge always makes things faster." Only if its data is near it. Edge code that makes several sequential calls to a distant database is slower than the same code at the origin.
Misconception
"Multi-cloud doubles availability." It protects only against failures confined to one provider, which are a small share of most services' downtime; it adds complexity that can cause new incidents. Multi-region within one provider captures most of the benefit for much less.
Misconception
"Our data is in a European region, so it is beyond the reach of foreign law." Residency is not sovereignty. A provider subject to another country's law may be compelled to disclose data it controls wherever it is stored; what limits that is who controls the keys and the operations, not only where the disks are.
Misconception
"GDPR requires EU data to stay in the EU." It does not require localisation; it requires a legal basis for transfers. Localisation requirements come from other rules — national or sectoral — such as India's for payment data or China's for critical infrastructure.
Misconception
"Only the primary database matters for residency." Backups, replicas, logs, analytics copies, caches and remote access by support staff all count, and are where most violations hide.
Exercises
1. The edge, multi-cloud and digital sovereignty
CoreHumanA user is 5 ms (round trip) from an edge site and 80 ms from the origin region; the edge site is also 80 ms from the origin. Each request opens a new connection (three round trips: TCP, TLS 1.3 and the request) and makes sequential queries to a database in the origin region; from a server beside the database a query takes 1 ms, from the edge 80 ms. Server computation is the same either way. What is the largest for which running the code at the edge is still faster than running it at the origin?
2. The edge, multi-cloud and digital sovereignty
CoreHumanAn edge function in Lagos handles each request by making three sequential queries to a database in Frankfurt (about 74 ms round trip each). Users find it slower than the old version that ran in Frankfurt. Which change helps most?
3. The edge, multi-cloud and digital sovereignty
CoreComputationalWrite greedy_sites(weights, sites, k, rtt) that chooses k edge sites greedily: start with none, and repeatedly add the site (from sites, not yet chosen) that gives the lowest weighted mean RTT, where each user goes to its nearest chosen site.
weightsmaps user name to number of users;rtt(user, site)returns a round-trip time.- The weighted mean RTT of a set is .
- Break ties by taking the site that comes first in
sites. - Return the list of chosen sites in the order chosen.
4. The edge, multi-cloud and digital sovereignty
CoreComputationalWrite violations(copies, policy) returning the names of the copies that break a residency policy, in their original order.
policymaps a data class to a pair(allowed_regions, safeguards);allowed_regionsis a set, orNonefor no restriction;safeguardsis a set of legal bases that make any other region acceptable.- Each copy is a tuple
(name, data_class, region, safeguard), wheresafeguardmay beNone. - A copy is fine if its region is allowed, or its safeguard is in the class's safeguards.
- A copy whose data class is not in the policy at all is a violation (deny by default).
5. The edge, multi-cloud and digital sovereignty
CoreHumanWhich statement about the EU's General Data Protection Regulation is correct, as of 2026?
6. The edge, multi-cloud and digital sovereignty
Warm-upHumanAn active-active system replicates 1.5 TB (1,500 GB) of changes per day from one provider to another, at an example egress price of USD 0.08 per GB. How many US dollars does replication cost in a 30-day month?
7. The edge, multi-cloud and digital sovereignty
CoreHumanA company must keep a class of personal data inside one country. Which of these must its data inventory include and check? List the letters.
- (a) the primary database
- (b) nightly backups copied to object storage
- (c) application logs that contain customers' email addresses
- (d) a support engineer in another country viewing customer records through an admin tool
- (e) the source code of the application
- (f) a cache at an edge site that stores rendered account pages
Further reading and resources
- RFC 8446: The Transport Layer Security (TLS) Protocol Version 1.3 · 2018
- RFC 9000: QUIC, a UDP-based multiplexed and secure transport · 2021
- Cloudflare Workers documentation
- Fastly Compute documentation
- Akamai EdgeWorkers documentation
- Amazon CloudFront Functions
- Google Cloud Service Extensions
- Alibaba Cloud ESA Edge Routine
- KubeEdgeA CNCF project extending Kubernetes to edge devices.
- Summary of the Amazon DynamoDB service disruption in the Northern Virginia (US-EAST-1) Region · Amazon Web Services · 2025
- Google Cloud incident report, 12 June 2025 · Google Cloud · 2025
- Cloudflare outage on November 18, 2025 · Cloudflare · 2025
- General Data Protection Regulation (EU) 2016/679
- Data Act, Regulation (EU) 2023/2854
- European Commission — adequacy decisions
- European Data Protection Board
- ANSSI (French national cybersecurity agency), home of SecNumCloud
- BSI (German Federal Office for Information Security), home of the C5 catalogue
- Ministry of Electronics and Information Technology, India (DPDP Act and Rules)
- Reserve Bank of India
- Cyberspace Administration of China
- Office of the Data Protection Commissioner, Kenya
- Autoridade Nacional de Proteção de Dados, Brazil
- US Department of Justice — CLOUD Act resources
- US Department of Justice — Data Security Program
- Opening the AWS European Sovereign Cloud · Amazon Web Services · 2026
- Microsoft Sovereign Cloud documentation
- Microsoft EU Data Boundary
- Google Cloud sovereign cloud
- Oracle sovereign cloud
- OVHcloud
- Scaleway
- STACKIT
- Gaia-XA European framework for trusted data sharing between providers, not itself a cloud.