Bulkheads: Partition and Tolerate Failure in One Part

Ideas you will meet

Resilience

Timeouts, retries, circuit breakers, bulkheads and load shedding.

What we already know about this

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

• Advanced Distributed Systems · 13 minResilience: timeouts, retries, bulkheads, backpressure and circuit breakersHow to stop one slow or failing dependency from taking everything down: the patterns, how they fit together and where each one goes wrong.Read hereClose

  The core problem

  In a system of many services, slowness is more dangerous than failure. A dependency that fails fast is easy to handle. One that responds in 30 seconds makes callers hold threads, connections and memory until they too stop responding: a cascading failure. Resilience patterns limit how far a problem can spread.

  The toolbox

  100%

  Figure SDA-ADS-06-1 · Resilience layers around a call · © forcsguy.com

  Timeouts

  Every network call needs a timeout, including connect, read and overall. Pick from measured latency (for example slightly above the 99th percentile), not a guess, and make the caller's budget smaller than its own caller's so work is abandoned in an orderly way. Propagate deadlines across calls.

  Retries

  Retry only transient failures and only idempotent operations (or use idempotency keys). Use exponential backoff with jitter (random spread) so clients do not retry in lockstep, cap the attempts, and apply a retry budget (for example no more than 10 percent extra traffic). Retry at one layer, not at every layer, or attempts multiply.

  Circuit breaker

  A breaker watches failures or slow calls. Closed: normal. After too many failures it opens: calls fail immediately (or use the fallback) so the dependency can recover and callers stop wasting resources. After a cool-down it goes half-open: a few trial calls decide whether to close again.

  100%

  Figure SDA-ADS-06-2 · Circuit breaker states · © forcsguy.com

  Bulkheads

  Partition resources so one dependency cannot consume them all: separate thread pools or connection pools per dependency, separate queues per tenant or per priority, or separate instances for critical paths. A slow recommendation service then cannot starve checkout.

  Backpressure and load shedding

  If producers outpace consumers, work piles up. Without limits the buffer grows until memory is exhausted or latency becomes huge. Options:

  • Bounded queues that reject or block when full.
  • Signalling upstream to slow down (credit-based flow in streaming systems, TCP windows, reactive streams).
  • Load shedding: deliberately drop low-priority work early (return 429 or 503 quickly) to protect the rest.
  • Rate limiting at the edge (see the rate-limiter lesson).

  Failing fast and visibly under overload is better than accepting everything and failing slowly.

  Fallbacks and graceful degradation

  Return something useful when a dependency fails: cached data, a default, a simpler algorithm, or hide a non-essential section. Decide which features are essential before an incident, and test the degraded mode.

  How they interact

  • A timeout is the minimum; without it nothing else works.
  • Retries inside a breaker inside a bulkhead is the usual nesting.
  • Everything needs metrics: breaker state, rejected calls, queue depth, timeout counts. Alert on them.
  • Test it: inject latency and failures (chaos experiments) to confirm the patterns work before production does it for you.

  Common mistakes

  Retries with no backoff; timeouts longer than the caller's; unbounded queues; a shared thread pool for everything; breakers that never close; fallbacks that call the same failing service; and testing only the happy path.

  Real-world reading

  Related reading

  • Timeouts, retries, and backoff with jitterAmazon Builders' Library · Amazon Web ServicesBlogs
  • Using load shedding to avoid overloadAmazon Builders' Library · Amazon Web ServicesBlogs
  • Circuit BreakerMartin FowlerBlogs
  • Circuit Breaker patternMicrosoft · Microsoft AzureDocumentation

  Show 6 moreShow fewer

  • Google SRE Book: Addressing Cascading FailuresGoogle SRE team · GoogleBlogs
  • Chaos Monkey (source)Netflix · GitHubGitHub
  • Bulkheads: Partition and Tolerate Failure in One PartSkife engineers · SkifeCase Studies
  • Protector: Circuit Breaker for Time Series Databases at TrivagoTrivago engineers · TrivagoCase Studies
  • Autoscaling Jenkins at TrivagoTrivago engineers · TrivagoCase Studies
  • Troubleshooting a Connection Timeout Issue with tcp_tw_recycle Enabled at eBayeBay engineers · eBayCase Studies

  Open this lesson with its quiz and exercise

• Core Concepts: How Things Actually Work · 9 minHow AWS Lambda and serverless platforms workWhat happens between an event and your function running: micro-VMs, cold starts, scaling and the limits that follow.Read hereClose

  What "serverless" means

  You provide a function; the platform decides where and when it runs, scales it up and down (even to zero), and bills for the time used. Servers still exist, but you do not manage them. AWS Lambda popularised the model, and Google Cloud Functions and Azure Functions work similarly.

  The path of an invocation

  100%

  Figure SDA-CORE-19-1 · Serverless invocation path · © forcsguy.com

  1. An event arrives from an API gateway, a queue, a file upload, a timer and so on.
  2. The platform's front-end service authenticates it and finds your function.
  3. A placement service looks for a free, warm execution environment for your function. If none exists it picks a worker host and creates one.
  4. The environment runs your code in an isolated micro-VM (AWS uses Firecracker, a lightweight virtual machine that starts in a fraction of a second) so that different customers' code is strongly isolated on shared hardware.
  5. The response flows back (synchronous calls), or the result is recorded (asynchronous calls).

  Cold and warm starts

  An execution environment goes through phases:

  1. Create the micro-VM and download your code or container image.
  2. Initialise the runtime and run your initialisation code (imports, creating clients).
  3. Invoke the handler.
  4. After finishing, the environment is frozen and kept for a while, ready to be reused (a warm start) for the next request.

  The first three steps together are a cold start. It adds tens to hundreds of milliseconds, or more for large packages and heavy runtimes. Warm requests skip straight to the handler.

  100%

  Figure SDA-CORE-19-2 · Synchronous and asynchronous invocation · © forcsguy.com

  Scaling

  The platform runs one request per environment at a time. More simultaneous requests mean more environments, created on demand, up to concurrency limits (account and per function). Scaling is fast but not instant, and a burst can hit cold starts. You can reserve capacity with provisioned concurrency, and cap a function so it cannot overwhelm a downstream database.

  Invocation types

  Type	Behaviour	Failure handling
  Synchronous (API calls)	The caller waits for the result	The caller sees the error and decides to retry
  Asynchronous (events)	Queued internally, run later	Automatic retries, then a dead-letter queue or failure destination
  Stream or queue polling	The platform reads batches from a queue or stream and invokes your function	Failed batches retry, so use partial-failure reporting and idempotent handlers

  Limits to design around

  • A maximum run time (for Lambda, 15 minutes at the time of writing), and limits on memory, package size and payload size. Memory settings also determine CPU share.
  • Ephemeral local storage: do not rely on it for state. Keep state in databases, caches or object storage.
  • No inbound connections to the function: it is invoked, it does not listen.
  • Concurrency pressure on dependencies: thousands of parallel functions can open thousands of database connections. Use a connection pooler or proxy, or limit concurrency.

  When serverless fits

  Good for spiky, event-driven or glue workloads, APIs with uneven traffic, scheduled jobs, file processing, and anything where idle time dominates. Less suited to very long-running jobs, steady high load (where always-on servers can be cheaper), latency-critical paths that cannot tolerate cold starts, or workloads needing special hardware.

  Practical tips

  • Keep packages small and imports lazy.
  • Do expensive setup outside the handler.
  • Make handlers idempotent and stateless.
  • Set timeouts and memory deliberately, and measure cost, since more memory can make a function faster and sometimes cheaper.
  • Add observability: structured logs, metrics for duration, errors and cold starts, and tracing across functions.

  Real-world reading

  Related reading

  • Using load shedding to avoid overloadAmazon Builders' Library · Amazon Web ServicesBlogs
  • GitHub Availability Report: June 2023GitHub Engineering · GitHubCase Studies
  • Avoiding insurmountable queue backlogsAmazon Builders' Library · Amazon Web ServicesBlogs
  • Timeouts, retries, and backoff with jitterAmazon Builders' Library · Amazon Web ServicesBlogs

  Show 6 moreShow fewer

  • Uber's Real-Time Push PlatformUber Engineering · UberCase Studies
  • NATS ServerNATS · CNCF / NATSGitHub
  • GitHub EngineeringGitHub engineers · GitHubBlogs
  • Bulkheads: Partition and Tolerate Failure in One PartSkife engineers · SkifeCase Studies
  • Data Platform at DoorDashDoorDash engineers · DoorDashCase Studies
  • API Platform at Riot GamesRiot Games engineers · Riot GamesCase Studies

  Open this lesson with its quiz and exercise

• Real-World Systems · 9 minUploading large files reliablyChunked, resumable, direct-to-storage uploads: how to move gigabytes over unreliable networks without tying up your servers.Read hereClose

  Why large uploads are hard

  A 5 GB file on a normal request has many ways to fail:

  • A single long connection dies on a flaky network, losing everything.
  • Timeouts in load balancers and proxies cut off long requests.
  • Your application servers are tied up carrying bytes, and memory or disk fills.
  • A failure at 99% means starting again.
  • Users want a progress bar and the ability to pause and resume.

  The design in one picture

  100%

  Figure SDA-RWS-17-1 · Chunked, resumable upload flow · © forcsguy.com

  Before: upload through your server

  100%

  After: upload straight to storage

  100%

  What changed

  • Your servers no longer carry large files
  • A failed chunk is retried, not the whole file
  • Pre-signed URLs keep storage private

  Step by step

  1. Initiate. The client tells the API the file name, size and type. The API checks permissions and limits, and creates an upload session with a unique id.
  2. Direct to storage. The API returns pre-signed URLs: short-lived, scoped links that let the client upload straight to object storage. Your servers never carry the file.
  3. Chunk. The client splits the file into parts (for example 5 to 100 MB; object stores typically accept thousands of parts per object) and uploads them, in parallel.
  4. Verify. Each part carries a checksum (MD5, CRC32C or SHA-256). The storage service or your server checks it, and a bad part is simply re-sent.
  5. Resume. The session remembers which parts are done. After a drop the client asks which parts exist and sends only the missing ones. Open protocols such as tus standardise resumable uploads.
  6. Complete. The client sends the list of parts; storage assembles the object.
  7. Post-process asynchronously. An event triggers virus scanning, type verification, thumbnails or transcoding, and metadata updates. The user sees "processing" until it finishes.

  Checking each part

  When a part arrives, the store replies with a checksum tag. The client keeps the list of part numbers and tags, and sends it with the final "complete" request, so the store can verify and assemble exactly the parts you meant, in order. A failed part is re-sent alone; assembly of a very large object can take a few minutes, so treat "complete" as asynchronous from the user's point of view.

  Smart extras

  • Deduplication: hash the file (or chunks) on the client first. If the content already exists in storage, skip the upload entirely and just add a reference (as in cloud file storage services).
  • Parallelism and adaptive chunk size: larger parts on fast networks, smaller on poor ones.
  • Compression for compressible data before upload.
  • Upload acceleration: send to the nearest edge location, which forwards over the provider's backbone.
  • Progress and cancellation: the client tracks bytes per part; cancelling aborts the session.

  Safety and cost

  • Limit size, type and number of concurrent uploads per user, and expire pre-signed URLs quickly.
  • Validate on the server, not only in the client: check file type by content, scan for malware, and store uploads in a quarantine area until cleared.
  • Use lifecycle rules to delete incomplete multipart uploads after a few days, since unfinished parts cost money until removed.
  • Encrypt in transit and at rest; do not let user-provided names choose storage paths.
  • Rate limit session creation to prevent abuse.

  Failure handling checklist

  Failure	Handling
  Connection drops mid-part	Retry that part with backoff
  Client closes the app	Session persists; resume later by listing stored parts
  Corrupted part	Checksum mismatch, re-upload that part
  Duplicate completion request	Idempotent: completing twice returns the same result
  Abandoned upload	Lifecycle rule cleans up the leftovers
  Processing fails	Retry from the event queue, mark the file as failed, tell the user

  Real-world reading

  Related reading

  • Determinism (4 parts) in League of Legends ServerRiotgames engineers · RiotgamesCase Studies
  • Implementing health checksAmazon Builders' Library · Amazon Web ServicesBlogs
  • The Network is ReliableKyle Kingsbury · aphyr.comBlogs
  • Bulkheads: Partition and Tolerate Failure in One PartSkife engineers · SkifeCase Studies

  Show 6 moreShow fewer

  • Natural Language Processing and Content Analysis (2 parts) at Condé NastCondé Nast engineers · Condé NastCase Studies
  • Autoscaling Bigtable Clusters based on CPU Load at SpotifySpotify engineers · SpotifyCase Studies
  • Ambry: Distributed Immutable Object Store at LinkedInLinkedIn engineers · LinkedInCase Studies
  • Big Data Processing (2 parts) at SpotifySpotify engineers · SpotifyCase Studies
  • Container (8 parts) at Riot GamesRiot Games engineers · Riot GamesCase Studies
  • Architecture of Chat Service (3 parts) at Riot GamesRiot Games engineers · Riot GamesCase Studies

  Open this lesson with its quiz and exercise