Cookie preferences

Essential cookies keep this site working. Optional Google Analytics cookies help us understand page visits and improve the site. You can change your choice anytime. Cookie policy

Skip to content
ForCSGuy System Design Atlas home

Library / Papers

Maglev: A Fast and Reliable Software Network Load Balancer (paper)

How Google runs its network load balancing in software on commodity servers, using a consistent-hashing variant to keep connections stable.

AdvancedLoad BalancerDistributed Systems

By Eisenbud et al. · Google · Added 2026-10-12

Read the paper

All content belongs to Eisenbud et al. / Google. We only link to it and give credit. Report a problem

Learn this paper

About 37 min left. Your work is saved as you go.

What you are about to study

Google wrote this paper. Read the abstract and headings first. It assumes you are comfortable with distributed systems.

Ideas you will meet

Load Balancer
Spread traffic across servers with L4/L7 balancing, health checks and algorithms.
Distributed Systems
Consensus, time, failure and the fundamentals of scale-out.

What we already know about this

Our own lessons on these ideas. Read them here before you start, so it makes sense.

  • System Design Foundations · 10 minFrom one server to many: scaling and load balancingVertical versus horizontal scaling, why stateless servers matter, and how a load balancer spreads traffic.Read here

    Two ways to grow

    When one server is not enough you have two options.

    A technician working at a server rack in a data centerImage: Derrick Coetzee · CC0 · Source · Wikimedia Commons: Technician with laptop working on server rack at NERSC

    • Vertical scaling ("scale up"): use a bigger machine with more CPU, memory or disk. It is simple because nothing in the design changes.
    • Horizontal scaling ("scale out"): add more machines and share the work between them. It needs more design effort but has no hard ceiling.
    VerticalHorizontal
    EffortLowHigher
    CeilingHardware limitPractically none
    FailureOne machine is a single point of failureSurvives losing a machine
    Cost curveGets steep for the largest machinesRoughly linear with commodity machines

    Real systems use both: scale up until it gets expensive or risky, then scale out.

    Figure SDA-FND-02-1 · Scale up versus scale out · © forcsguy.com

    What a load balancer does

    Once you have several servers, something must decide which one handles each request. That is the load balancer.

    Figure SDA-FND-02-2 · Load balancer in front of a server pool · © forcsguy.com

    It does three jobs:

    1. Distributes requests so no server is overloaded.
    2. Checks health. It probes each server regularly and stops sending traffic to ones that fail.
    3. Hides the pool. Clients talk to one address, and servers can be added or removed without clients noticing.

    Common strategies

    • Round robin: take turns. Simple and fine when requests are similar.
    • Least connections: send to the server with the fewest active requests. Better when request cost varies.
    • Hashing (for example on user id or IP): the same key goes to the same server. Useful when a server keeps a local cache for that key.
    • Weighted: bigger machines get a bigger share.

    Load balancers work at different layers. A layer 4 balancer routes by IP and port and is very fast. A layer 7 balancer reads the HTTP request, so it can route by URL path or header and terminate TLS.

    A security benefit comes with this layout: only the balancer needs a public address. The servers behind it use private addresses, reachable from inside the network but not from the internet, so clients cannot bypass the balancer or attack a server directly. If one server goes offline, the balancer sends its traffic to the others and a replacement can join the pool.

    Before: one server

    After: load balancer and several servers

    What changed

    • Traffic is shared across servers
    • One server failing no longer takes the site down
    • You can add capacity without downtime

    Stateless servers make scaling easy

    A server is stateless when it keeps nothing about a user between requests. Everything it needs comes from the request or from a shared store.

    Figure SDA-FND-02-3 · Stateful versus stateless servers · © forcsguy.com

    With stateless servers any request can go to any server, so you can add servers at will, replace a failed one without losing data, and run auto-scaling. The state does not disappear. It moves to a database, a cache or a session store, which you then have to scale and protect, which is what the next lessons are about.

    Putting it together

    A typical first scalable layout:

    1. Users reach a load balancer.
    2. It spreads requests over a pool of stateless application servers.
    3. Servers read and write a shared database (and later a cache).
    4. Static files such as images and scripts are served from a CDN so the servers do not carry that traffic (lesson 9).

    This design handles a surprisingly large amount of traffic. The first thing to give way is usually the database, which is where we go next.

    A worked example: how many servers do you need?

    Capacity planning is arithmetic you can do on a napkin. Say your product has 5 million daily users and each makes about 20 requests a day.

    1. Requests per day: 5,000,000 x 20 = 100 million.
    2. Per second on average: 100,000,000 / 86,400 = about 1,160 requests per second.
    3. Traffic is not flat. A common rule of thumb is that the busiest second is 2 to 3 times the average, so plan for about 3,500 per second.
    4. One server handles about 400 requests per second flat out, but you never want servers at 100%: queues build and latency explodes near full load. Plan for roughly 70% utilisation, so 280 requests per second each.
    5. 3,500 / 280 = 12.5, so 13 servers carry the peak.
    6. Add spare capacity so you survive failures: with 13 servers, losing 2 leaves 11, which carry only 3,080 requests per second. Plan for N+2: 15 servers.

    Figure SDA-FND-02-4 · From users to a server count · © forcsguy.com

    The same method works for any tier. Only the "requests per server" number changes, and you should measure it with a load test rather than guess.

    Choosing a balancing method with real numbers

    Round robin is fair in the number of requests, not in the work. Imagine three servers and a stream where most requests take 10 ms but every fifth takes 1 second (a report export, say). Round robin gives each server the same count, but by chance one server collects several slow requests at once, its queue grows, and its users wait while the other two are nearly idle.

    Least connections notices: the server with the most open requests gets fewer new ones, so the slow requests spread out as they arrive. It costs the balancer a little bookkeeping and helps most when request costs vary a lot. Use plain round robin when requests are similar and servers identical; use least connections, or "power of two choices" (pick two servers at random and send to the less busy one), when they are not. For caches and stateful workloads, consistent hashing keeps the same key on the same server so its memory stays useful.

    Health checks and draining

    A balancer is only as good as its knowledge of which servers are alive.

    • Active health checks call a lightweight endpoint (for example /health) every few seconds. After two or three failures in a row, the server leaves rotation; after a few successes it returns. Make the check reflect real readiness: a server that is up but cannot reach its database should fail the check.
    • Passive checks watch real traffic: too many errors or timeouts from one server and it is ejected.
    • Connection draining matters at deploys. Before stopping a server, tell the balancer to send it no new requests, let in-flight requests finish (typically 30 to 60 seconds), then stop it. Without draining, every deploy drops a few user requests.

    Figure SDA-FND-02-5 · Rolling out a new version without dropping requests · © forcsguy.com

    Mistakes people make

    • Treating the balancer as magic. It is a server too and can be a single point of failure. Run at least two, or use a managed one.
    • Adding servers behind a single database. App servers scale out easily; if the database is the bottleneck, more app servers only make it fail faster.
    • Keeping state on the server. Sessions, uploads and caches held in one server's memory break the moment a request lands elsewhere. Move them to a shared store.
    • Sizing for the average. The system fails at peak, which is when it matters.
    • Never testing failure. Kill a server during a load test. If users notice, you have found a gap before they did.

    Try it yourself

    Open the System Builder challenge "One server is not enough". Run the test on one server and watch it fail, add a load balancer without adding servers, then add servers and see where the numbers settle. Or skip the challenge and open a load-balanced system on the free canvas, then change the wiring yourself.

    More on this site

    Related reading

    Show 6 more

    Open this lesson with its quiz and exercise

  • Advanced Distributed Systems · 12 minLoad balancing: layer 4, layer 7 and gatewaysWhat a load balancer can see and do at each layer, the algorithms behind it, and how API gateways add routing, aggregation and policy.Read here

    What a load balancer sees

    A load balancer spreads incoming work across servers, removes failed servers and makes scaling invisible to clients. What it can do depends on which layer it works at.

    Layer 4 (transport)Layer 7 (application)
    SeesIP addresses, ports, TCP/UDP connectionsHTTP method, URL, headers, cookies, body
    Decides byConnection detailsRequest content
    TLSPasses through, or terminates in some productsUsually terminates TLS
    Speed and costVery fast, low overheadMore work per request
    Typical featuresSpread connections, health checksPath routing, header rules, sticky sessions, rewrites, caching, WAF rules
    ExamplesNetwork load balancers, LVS, HAProxy in TCP modeNGINX, Envoy, HAProxy HTTP mode, cloud application balancers

    Figure SDA-ADS-05-1 · Layers of balancing in front of an app · © forcsguy.com

    Many designs use both: an L4 tier for raw scale and an L7 tier for smart routing.

    Before: layer 4: ports and addresses

    After: layer 7: reads the request

    What changed

    • Layer 7 can route by path, header or cookie
    • Layer 7 can terminate TLS and retry safely
    • Layer 4 is simpler and faster, with less to configure

    Choosing a target: algorithms

    • Round robin / weighted round robin: simple rotation; weights favour bigger machines.
    • Least connections (and weighted): send to the server with fewest active requests; good when requests vary in length.
    • Least response time: combine active load with measured latency.
    • IP hash / consistent hashing: the same client or key goes to the same server, useful for caches and sticky state (see consistent hashing).
    • Resource-based: servers report load and the balancer follows.

    Health checks (active probes and passive error tracking) remove unhealthy servers; connection draining lets in-flight requests finish before a server is removed during a deploy.

    Global and DNS balancing

    Across regions, DNS-based or global server load balancing directs users to the nearest or healthiest region (see the DNS and CDN lessons). DNS answers are cached, so failover takes as long as the TTL; anycast and global balancers react faster.

    API gateway: policy at the edge

    An API gateway is an L7 component specialised for APIs. Beyond routing it commonly provides:

    • Authentication and authorisation (validate tokens, apply scopes),
    • Rate limiting and quotas (see the rate-limiter lesson),
    • Request and response transformation, versioning, and protocol translation,
    • Aggregation: one client call fans out to several services and merges results, saving round trips for mobile clients,
    • Caching, logging, metrics and tracing in one place.

    Figure SDA-ADS-05-2 · Gateway routing and aggregation · © forcsguy.com

    Risks: the gateway is a single entry point (run several instances, keep it stateless) and can become a bottleneck or a place where business logic piles up. Keep it to cross-cutting concerns. A variation is a backend for frontend: one gateway per client type (web, mobile) tailored to its needs.

    Inside the cluster

    Between services, balancing is done by client-side libraries or sidecar proxies that know the healthy instances from service discovery, so no central balancer is needed for east-west traffic. See the microservices lessons for service mesh ideas.

    Failure habits

    • Run at least two of every balancer and gateway instance across zones.
    • Keep timeouts and retry budgets (see the resilience lesson) so the balancer does not amplify an outage.
    • Use slow start for new instances so they warm up before taking a full share.

    Real-world reading

    Related reading

    Show 6 more

    Open this lesson with its quiz and exercise

  • System Design Foundations · 12 minDesigning for failure: reliabilityTimeouts, retries, circuit breakers, rate limiting and redundancy: how systems stay up when parts of them break.Read here

    Failure is normal

    At scale something is always broken: a disk, a network link, a deployment, a dependency having a bad day. Reliable systems do not avoid failure, they expect it and contain it.

    Two words to keep apart:

    • Availability: the fraction of time the system works (99.9% allows about 8.8 hours of downtime a year; 99.99% about 53 minutes).
    • Durability: the chance stored data is not lost.

    Remove single points of failure

    A single point of failure (SPOF) is any component whose failure takes everything down. Hunt for them:

    • Run several application servers behind a load balancer.
    • Replicate databases and plan failover (lesson 5).
    • Spread across availability zones, and for the most critical services, across regions.
    • Do not forget the "invisible" ones: the load balancer, DNS, a certificate that expires, a single queue.

    Figure SDA-FND-10-1 · Removing single points of failure tier by tier · © forcsguy.com

    Protect calls between services

    Most outages spread through calls between components. Four basic defences:

    1. Timeouts

    Every network call needs a timeout. Without it a slow dependency holds your threads hostage. Set timeouts based on what a healthy call looks like (for example a little above the 99th percentile), and make inner timeouts shorter than outer ones.

    2. Retries, carefully

    Transient errors often succeed on a second try. But retries multiply load on a struggling service, so:

    • Retry only idempotent operations or those with an idempotency key (lesson 8).
    • Use exponential backoff with jitter.
    • Cap the number of attempts, and ideally cap the share of traffic that can be retries.

    3. Circuit breakers

    If a dependency keeps failing, stop calling it for a short time and fail fast.

    A vintage electrical circuit breaker, the physical idea behind the software circuit breakerImage: Alf van Beem · CC0 · Source · Wikimedia Commons: Circuit breaker, Single pole, Non-tripfree, 20Amp-30Volts, AN 3160 pic1

    Figure SDA-FND-10-2 · Circuit breaker states · © forcsguy.com

    • Closed: normal operation, failures are counted.
    • Open: calls fail immediately, saving resources.
    • Half-open: a few trial calls check whether the dependency has recovered.

    4. Bulkheads

    Give each dependency its own pool of threads or connections, like watertight compartments in a ship, so one slow dependency cannot use up everything.

    Before: a slow dependency drags everyone down

    After: a breaker cuts the slow dependency off

    What changed

    • Callers stop waiting on a failing service
    • Threads and connections are not used up
    • A fallback answer is returned while the service recovers

    Protect yourself from too much traffic

    • Rate limiting caps how many requests a client may make (common algorithms: token bucket, sliding window). It protects against abuse and accidental floods.
    • Load shedding: when overloaded, deliberately reject low-priority work early so important requests still succeed. It is better to serve 80% of requests well than 100% badly.
    • Back pressure: tell producers to slow down instead of letting queues grow without bound.

    Degrade gracefully

    When a non-critical part fails, keep the rest working: show the page without recommendations, serve cached prices, queue the email. Decide in advance which features are essential.

    Practise failure

    • Health checks so traffic leaves broken instances.
    • Chaos testing: deliberately kill instances in a controlled way to prove the system copes.
    • Runbooks and game days so people have rehearsed the response.

    A worked example: the arithmetic of availability

    Downtime in a month. 99% allows about 7.3 hours; 99.9% about 43 minutes; 99.99% about 4.4 minutes; 99.999% about 26 seconds. Each extra nine is ten times harder and usually much more expensive, so choose the target your users need, not the biggest number.

    Dependencies in series multiply. If a request needs three services, each 99.9% available, the request is available only 0.999 x 0.999 x 0.999 = 99.7%, about 2 hours 11 minutes of downtime a month. The more things in the critical path, the worse it gets, which is why teams shorten the critical path and make non-essential calls optional.

    Redundancy in parallel helps. Two independent replicas, each 99% available, are both down only 1% x 1% = 0.01% of the time: 99.99% together. The word "independent" carries all the weight: replicas that share a rack, a deploy, a bug or a dependency fail together, and the formula no longer applies.

    Retries multiply load. If every one of 1,000 requests per second fails and each caller retries 3 times, the failing service sees up to 4,000 requests per second, making recovery harder. And if three layers each retry 3 times, one user request can become 64 attempts at the bottom. Retries need limits, backoff with jitter, and a retry budget (for example, retries may be at most 10% of normal traffic).

    Figure SDA-FND-10-3 · How availability combines · © forcsguy.com

    Anatomy of a cascading failure

    The classic outage is rarely one failure; it is one failure that spreads.

    1. A database gets slow (a bad query, a noisy neighbour). Responses that took 10 ms now take 5 seconds.
    2. App servers wait on it. Each request holds a thread for 5 seconds, so threads and connections fill up.
    3. The app can no longer accept new requests, even ones that do not need the database. The load balancer sees timeouts and marks servers unhealthy, so the same traffic lands on fewer servers.
    4. Callers time out and retry, which adds load to the already struggling system.
    5. Everything is down, and fixing the original slow query no longer brings it back, because the retry storm keeps it overloaded.

    The defences are cheap compared with the outage:

    • Timeouts on every call, shorter than the caller's own timeout.
    • Circuit breakers that stop calling a failing dependency for a while and fail fast.
    • Bulkheads: separate thread or connection pools per dependency, so one slow one cannot use all of them.
    • Load shedding: when overloaded, reject some requests early to keep serving the rest.
    • Backoff with jitter and retry budgets.

    Figure SDA-FND-10-4 · A circuit breaker protects both sides · © forcsguy.com

    Mistakes people make

    • No timeouts, or timeouts longer than the caller's. The default in many libraries is "wait forever" or minutes.
    • Retrying everything everywhere. Retry only safe, idempotent operations, at one layer, with backoff, jitter and a budget.
    • Health checks that lie. A server that answers /health but cannot reach its database should fail the check.
    • Untested failover. Failover that has never run is a guess. Practise it on purpose, in business hours.
    • Shared fate. Two "redundant" systems in one zone, on one deploy pipeline, or on one dependency are one system.
    • No graceful degradation. Decide in advance which features can be switched off (recommendations, images, analytics) to protect the critical path.

    Try it yourself

    In the System Builder, lose three servers at peak and find how much spare capacity you need. Read the resilience patterns lesson for circuit breakers and bulkheads in more depth, and build the lease leader election lab to see failover without split brain.

    More on this site

    Related reading

    Show 6 more

    Open this lesson with its quiz and exercise

Continue to Google

Go deeper on this

A path through videos, articles, other case studies and code on the same ideas as Maglev: A Fast and Reliable Software Network Load Balancer (paper).

  1. Warm upGet the vocabulary first.

  2. WatchHow others solve the same problem, on video.

    Videos belong to their creators and play from YouTube. More on YouTube (opens in a new tab)

  3. ReadEngineers explain it in their own words.

  4. See it in another companyThe same ideas, a different system.

  5. Read the codeSee how it looks when it is real.

Keep learning

Loading community…