Microservices

In simple terms

A microservice architecture breaks an application up into many small services that each do one thing — auth, billing, search, notifications — and talk to each other over the network. The opposite is a monolith, where one big codebase and one process does everything.

The Visual Map

flowchart LR
  subgraph mono["monolith"]
    M["one process:<br/>auth + billing + search<br/>+ notifications"] --> MD[("one database")]
  end
  subgraph micro["microservices"]
    A["auth svc"] --> AD[("auth DB")]
    B["billing svc"] --> BD[("billing DB")]
    S["search svc"] --> SD[("search index")]
    A -.HTTP/gRPC.-> B
    B -.events via queue.-> N["notify svc"]
  end
  mono -->|"in-process call becomes<br/>a network call"| micro

More detail

Often-cited benefits:

Independent deploys — each team ships their service on their own cadence.
Independent scaling — scale the busy service, not everything.
Technology diversity — different services can pick their own language and database.
Bounded fault impact — a failure in one service doesn’t crash the whole app, if the rest is built to degrade gracefully.
Smaller teams own smaller code — Conway’s law: org and architecture mirror each other.

Costs and risks (always understated):

Distributed systems are harder — network failures, retries, partial outages, eventual consistency.
Operational complexity — many services means many deploys, dashboards, on-call rotations.
End-to-end debugging — a single user request can traverse a dozen services.
Data consistency — no cross-service transactions; you live with sagas, outbox patterns, and reconciliation jobs.

Rules of thumb:

Start with a monolith. Split out a service only when a clear seam exists and the cost is justified.
A service that just wraps a database table is too small.
A service owned by no team is doomed.

Supporting tech you usually adopt with microservices: containers, an orchestrator (Kubernetes), service discovery, distributed tracing, centralised logging, a CI/CD pipeline that can ship dozens of services.

Microservices, when right-sized, let large organisations move quickly without coordinating every change across the whole company. When wrong-sized, they recreate every problem a monolith had plus the entire surface area of a distributed system.

Under the Hood

The architecture, as deployment config — each service independently built, versioned, scaled, and owned:

# docker-compose.yml — a microservice system in miniature
services:
  auth:
    image: registry.local/auth:2.4.1      # its own version...
    environment: { DB_URL: postgres://auth-db/auth }
    deploy: { replicas: 2 }

  billing:
    image: registry.local/billing:7.0.3   # ...on its own release cadence
    environment:
      DB_URL: postgres://billing-db/billing
      AUTH_URL: http://auth:8080          # cross-service call = network call
    deploy: { replicas: 4 }               # scaled independently

  notifications:
    image: registry.local/notify:1.9.0
    environment: { QUEUE_URL: amqp://broker }   # async via the queue

  auth-db:    { image: postgres:17 }      # one database PER service —
  billing-db: { image: postgres:17 }      # no shared tables, no cross-DB joins

The two load-bearing decisions are visible right here: every call between services crosses the network (so it can fail), and every service owns its data exclusively (so there are no cross-service transactions).

Engineering Trade-offs

Team autonomy vs systemic complexity. Independent deploys remove cross-team coordination — and replace it with versioned APIs, backwards-compatibility contracts, and integration environments. You trade meeting overhead for engineering overhead.
In-process calls become network calls. A function call that couldn’t fail now has latency, timeouts, retries, and partial failure. Tail latency compounds: a request fanning out to 10 services at p99=100ms has a much worse p99 itself.
Per-service databases vs transactions. Data ownership enables independent evolution, but “debit billing, credit wallet” is no longer one ACID transaction — it’s a saga with compensation logic and reconciliation jobs.
Platform prerequisite. The architecture only pays off on top of mature CI/CD, observability, and orchestration. Adopting microservices before that platform exists is how teams end up slower than their old monolith — the lesson behind Uber’s consolidation of ~2,200 services.

Real-world examples

Amazon’s “two-pizza team” rule famously aligns team size with service ownership.
Netflix is the canonical microservices case study.
Shopify ran a monolith for years and proudly still mostly does — successfully.
Uber publicly walked back parts of its microservices architecture in 2020, consolidating ~2,200 services into larger ones because the operational overhead exceeded the agility benefit.

Common misconceptions

“Microservices are how big tech got to scale.” Some did; others scaled monoliths for many years. Architecture is one input among many.
“Microservices ship faster.” They can — but only if the platform underneath them is mature. Without it, they often slow teams down.

Try it yourself

Run two “services” and watch an in-process call become a fallible network hop — including what happens when the dependency is down:

python3 -c "
import threading, urllib.request, urllib.error
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer

class Auth(BaseHTTPRequestHandler):
    def do_GET(self):
        self.send_response(200); self.end_headers()
        self.wfile.write(b'{\"user\": \"ada\", \"valid\": true}')
    def log_message(self, *a): pass

auth = ThreadingHTTPServer(('127.0.0.1', 0), Auth)
threading.Thread(target=auth.serve_forever, daemon=True).start()
auth_url = f'http://127.0.0.1:{auth.server_address[1]}'

# billing service calling auth service:
print('auth up  :', urllib.request.urlopen(auth_url, timeout=1).read().decode())
auth.shutdown(); auth.server_close()
try:
    urllib.request.urlopen(auth_url, timeout=1)
except (urllib.error.URLError, ConnectionError) as e:
    print('auth down: billing must now decide — fail? retry? degrade?')
"

That final question is the daily texture of microservice engineering; circuit breakers and fallbacks are the standard answers.

Learn next

Container — the packaging unit that made this architecture practical.
Message queue — the async alternative to fragile call chains.
Service mesh — retries, mTLS, and routing factored out of every service.
Saga pattern — transactions when your data spans services.