Layer 4 and layer 7 load balancing

A layer 4 balancer forwards connections without looking inside, and a layer 7 balancer reads the request and decides with what it finds.

Use it when#

L7: you want to route by path, header, or kind of request, or retry a failed call on another server.
L4: raw throughput and low latency matter more than routing, or the traffic is not HTTP.
Either: a single balancer is the busiest node on the board.

How it works#

An L4 balancer works on IP addresses and ports. It picks a backend and passes the bytes along, so it does little work per request and handles very large volumes. An L7 balancer terminates the connection, parses HTTP, and can send /api to one pool and /images to another, add headers, and retry. Real examples are AWS's Network Load Balancer for L4 and its Application Load Balancer for L7.

Splitting pools by kind also isolates them: a flood of page views cannot starve the API servers.

What it costs#

L7 does more work per request, so it has a lower ceiling, costs more per request, and adds latency. L4 cannot see paths or retry a request, and it spreads load by connection, so one long-lived connection can pin a server.

In Quorum#

A Load Balancer has a variant. L7 is the default, and L4 needs L4 load balancing. L4 forwards every kind of request on every Link, so per-kind routing is gone. On an L7 balancer, each Link can carry only some kinds of request.

Drag to pan. Hold Ctrl or Cmd and scroll to zoom.

At 4.00k Users the Simulation solves this Design to Apdex 1.00, p99 300 ms, 0.0% errors and $1.52/s Burn.

Where	Setting	Value	Why
Load Balancer	Variant	l7	An L7 balancer reads the HTTP request, so it can send a kind of request to the servers built for it.
App Server	Instance size	small	Small gives the app two vCPUs, enough to run code while the data tier does its own work.
App Server	Instance size	small	Small gives the app two vCPUs, enough to run code while the data tier does its own work.
Database	Instance size	small	Small gives the Database two vCPUs, so the figures below reflect the Cache and not a starved Database.
Link Load Balancer → App Server	Kinds on this Link	page	Only page views go to this server, so its CPU and memory serve one workload.
Link Load Balancer → App Server	Kinds on this Link	read,write,login	API calls go here, so a burst of page views cannot starve them.

Research this Design needs: Load balancing.

Drag to pan. Hold Ctrl or Cmd and scroll to zoom.

At 4.00k Users the Simulation solves this Design to Apdex 1.00, p99 298 ms, 0.0% errors and $1.39/s Burn.

Where	Setting	Value	Why
Load Balancer	Variant	l4	An L4 balancer forwards TCP without reading HTTP, so it handles far more traffic for less money. It cannot route by kind or retry.
App Server	Instance size	small	Small gives the app two vCPUs, enough to run code while the data tier does its own work.
App Server	Instance size	small	Small gives the app two vCPUs, enough to run code while the data tier does its own work.
Database	Instance size	small	Small gives the Database two vCPUs, so the figures below reflect the Cache and not a starved Database.

Research this Design needs: Load balancing, Load balancing algorithms, L4 load balancing.

The glossary defines load balancer. Pair this with Load balancing algorithms.