Skip to content

Stateless tiers behind a load balancer

Separate the routing unit, the eligible targets and the owner of session state.

Predict first

Before you read: predict

Two fresh proxies each have healthy targets A and B and round-robin selection starting at A. One is L4, the other L7 without affinity. Each receives six sequential HTTP/1.1 reads over one downstream keep-alive connection.

What are their request totals, written A:B?

Pick one

The 30-second version

Horizontal scaling adds service instances behind a balancer. L4 selects a target for a connection; L7 can select per HTTP request. Keep-alive does not make those equivalent. Health checks change eligible targets. Stickiness retains a session's target while valid, concentrating busy sessions. Keep authoritative session state outside replaceable instances so reassignment does not restart progress. [S1–S4]

Separate the routing unit, the eligible targets and the owner of session state.

How it works

Step 1 of 8

Add replicas. Each instance can serve the same interface; adding one does not automatically redistribute existing connections. [S1]

Analogy: A hair salon's stations

A salon host can assign an entire appointment to one station, or assign each separate service visit at the desk. A return-to-the-same-stylist preference keeps one customer's visits together. A closed station disappears from the host's availability board, but that action does not copy its private notes. A shared service record lets the next station continue the agreed work. This is a simplified routing analogy, not guidance about hair care.

Salon element System meaning
Whole appointment assigned once L4 connection target
Each service visit assigned at the desk L7 request selection
Same-stylist preference Session affinity
Station availability board Health membership
Shared service record State outside replaceable instances

An appointment is not an HTTP stream: people can change stations mid-visit. The useful correspondence is when the host makes a routing decision. Nor does preference promise equal workloads: one customer can request many services while another requests only one. Our local request ledger makes this difference countable. [E1]

Try it

Answer first, then check. Nothing is saved.

Question 1 of 5

A catalog proxy is configured with layer=L7, sticky=true, two healthy targets, and round-robin selection for a new session. Requests arrive in this order: hot, cold, hot, hot, third, hot, fourth, hot, hot, hot.

What is the total A:B?

Pick one

Question 2 of 5

Fresh proxies use layer=L4 and layer=L7 respectively. One read is sent before A fails its health check while B remains healthy; both health rounds finish.

What happens to the next read through L4, then L7?

Pick one

Question 3 of 5

A proxy begins with sticky=true, only A eligible, and a session ledger containing returning → A. B then becomes eligible. It receives three reads with session returning, followed by one read each for new-one and new-two.

What are the A:B totals for the returning group, then the new group?

Pick one

Question 4 of 5

A session writes cursor 9 on A, is reassigned after A is removed, then A returns before four later reads. The candidate stores its cursor in the shared cursor service and selects a healthy target for every HTTP request (sticky=false).

Which [cursor,A-count,B-count] tuple does that candidate record?

Pick one

Question 5 of 5

Two sequential reads through a non-sticky L7 catalog proxy return one 503 and one 200. Both targets passed the health round that completed just before the reads.

Which defect best explains the symptom?

Pick one

Go deeper

Pro tip

Log the selected target identifier alongside a request's session identifier in a synthetic test. Aggregate views can hide whether a skew follows one busy session or many connections. Keep those test identifiers separate from real user credentials or personal data. [E1]

Sources

All fetched/read 2026-09-29. S references support platform rules; E1 supplies this fixture's observations and arithmetic.

Related lessons

Checked against the tools and docs as of 2026-09-29.