CAP in Practice: What AP Actually Buys You
Network partitions are not rare events, and the interesting systems never pick one side cleanly.
2 min readdistributed-systems
Archive
Every entry
103 long-form notes, newest first. No newsletter, no tracking — just the archive and an RSS feed.
Topics
- algorithms
- api-design
- architecture
- auth
- caching
- consensus
- consistency
- cryptography
- databases
- distributed-systems
- dns
- filesystems
- hashing
- infrastructure
- messaging
- networking
- observability
- performance
- pki
- post-quantum
- privacy
- reliability
- security
- storage
- testing
- tls
- web
- zero-knowledge
2026
42 entriesSymmetric vs Asymmetric Crypto: When to Use Which
One key to share, two keys to manage. A decision guide with the protocols that mix both.
2 min readcryptography
What Hash Functions Actually Guarantee
Avalanche, collision resistance, preimage resistance — the three properties that make SHA-256 boring on purpose.
2 min readhashing
Merkle Trees Beyond Blockchains
The hash tree is not a blockchain primitive — it is a general proof-of-membership machine, and most systems use it badly.
1 min readhashing · distributed-systems
Keyed vs Unkeyed Hashes
SHA-256 and HMAC both output 32 bytes, and using the wrong one in the wrong place is a real vulnerability class.
1 min readhashing · security
Hash Tables vs Cryptographic Hashes
Same word, opposite requirements: one optimizes for speed and spread, the other for unpredictability.
1 min readhashing · algorithms
The Birthday Bound in Practice
Why 128-bit hashes are enough for most things and 64-bit hashes are not enough for anything adversarial.
1 min readhashing · security
Salting and Peppering Password Hashes
A salt defeats precomputation; a pepper defeats a stolen database. Neither replaces a slow hash.
1 min readhashing · auth
Choosing Argon2 Parameters
Memory, iterations, and parallelism are a budget, not a magic tuple — here is how to spend it.
1 min readhashing · auth · performance
Content Addressing Is a Cache Key and a Contract
Hashing bytes to name them gives you free deduplication, immutable caching, and a verification step — if you keep the name honest.
1 min readhashing · storage
Bloom Filters and the Cost of False Positives
A bloom filter never lies about absence, but its false-positive rate is a tunable bill you pay in bits and hashes.
1 min readhashing · algorithms
Consistent Hashing Without the Hand-Waving
Virtual nodes, and why the ring is really a weighted random assignment that happens to be cheap to update.
1 min readhashing · distributed-systems
Hash Flooding and Why SipHash Exists
If an attacker can choose your map keys, a deterministic hash function is a denial-of-service primitive.
1 min readhashing · security
AEAD Nonce Reuse Is Catastrophic, Not Unlucky
Reusing a nonce under the same key leaks plaintext XOR and, for GCM, the authentication key itself.
1 min readcryptography · security
ChaCha20-Poly1305 vs AES-GCM
Both are safe AEADs. The choice is about hardware, side channels, and nonce size — not raw strength.
1 min readcryptography · performance · security
HKDF: Turning One Secret Into Many Keys
Extract-then-expand is the standard way to derive purpose-bound keys from shared secrets. Skip it and you invite cross-protocol attacks.
1 min readcryptography · tls
ECDSA Signatures Are Malleable
For every valid ECDSA signature there is a second valid one. Systems that assume uniqueness get burned.
1 min readcryptography · security
What Ed25519 Gets Right
Deterministic nonces, a cofactor-safe verification equation, and small keys — the modern default for signatures.
1 min readcryptography · security
X25519 and the Shared Secret You Never Use Directly
The raw Diffie-Hellman output is not a key. Feed it through a KDF and bind the transcript.
1 min readcryptography · tls
RSA Without Proper Padding Is Broken
Textbook RSA is multiplicative; PKCS#1 v1.5 is fragile; OAEP is the floor for encryption.
1 min readcryptography · security
Forward Secrecy Is About Key Deletion, Not Math
A session stays secret after a server compromise only if the ephemeral private key is actually gone.
1 min readcryptography · tls
Writing Constant-Time Code Without Lying to Yourself
Branching on secret bytes leaks through timing. Here is what actually helps and what only looks like it does.
1 min readcryptography · security
Randomness Failures Are Silent Until They Are Fatal
Weak RNGs do not throw errors. They produce keys that look fine and are shared with everyone.
1 min readcryptography · security
TLS 1.3 in One Round Trip
What was removed from TLS 1.2 matters as much as what was added: fewer options, fewer downgrade paths.
1 min readtls · cryptography
Certificate Transparency Makes Mis-Issuance Visible
CT does not stop a bad certificate; it guarantees someone can find it.
1 min readtls · pki · security
Revocation Is Still the Weakest Link
CRLs are too big, OCSP leaks browsing, stapling helps until it does not. Plan for revocation failure explicitly.
1 min readtls · pki
HPKE: Hybrid Encryption as a Standard Boring Primitive
One RFC that packages KEM, KDF, and AEAD so you stop inventing your own sealed-box format.
1 min readcryptography · tls
Mutual TLS Without the Operational Surprise
Client certificates are a strong authentication mechanism and a lifecycle problem. Treat them like credentials, not config.
1 min readtls · auth · security
SNI Leaks Every Domain You Visit
Encrypted Client Hello closes the last plaintext metadata hole in TLS — where it is deployed.
1 min readtls · privacy
Certificate Pinning Is a Trade-Off, Not a Free Win
Pinning stops MITM proxies and breaks your app when you forget to rotate the pin.
1 min readtls · security
ALPN: Negotiating Protocols Inside TLS
One port, many protocols. ALPN is how HTTP/2, HTTP/3, and custom RPCs coexist without a second handshake.
1 min readtls · networking
QUIC Folds Transport and Crypto Into One Handshake
TLS 1.3 runs inside QUIC, so the first flight already carries application data — and loses TCP head-of-line blocking.
1 min readnetworking · tls · performance
DNSSEC and DANE: Trust Anchors Without a CA
DNSSEC authenticates DNS answers; DANE can bind certificates to them. Both are powerful and operationally demanding.
1 min readdns · tls · pki
Linearizability vs Serializability
One is about real-time order of single operations, the other about equivalent serial execution of transactions.
1 min readdistributed-systems · consistency
Raft in One Page
Leader election, log replication, and safety — the three mechanisms that make Raft teachable.
1 min readdistributed-systems · consensus
Paxos vs Raft: Same Guarantees, Different Ergonomics
Paxos is a family of protocols; Raft is a specification with an implementation manual. Choose by ecosystem, not theory.
1 min readdistributed-systems · consensus
Quorum Math: Why Odd Numbers and Majority Rules
R + W > N gives consistency; majority quorums give availability under f failures. The rest is arithmetic.
1 min readdistributed-systems · consensus
Leader Leases Turn Consensus Into a Cache
A time-bounded lease lets a leader serve reads without a round trip — as long as clocks behave.
1 min readdistributed-systems · consensus
Clock Skew and TrueTime
Physical clocks are wrong by an unknown amount. Spanner turns that uncertainty into a wait.
1 min readdistributed-systems · consistency
Hybrid Logical Clocks Keep Causality and Wall Time
HLC timestamps stay close to physical time while never violating happens-before.
1 min readdistributed-systems · consistency
CRDTs: Convergence Without Coordination
If every replica applies operations commutatively, order stops mattering. The cost is metadata and semantics.
1 min readdistributed-systems · consistency
Gossip Protocols Scale by Being Sloppy
Random fanout spreads state in O(log n) rounds and never needs a membership oracle.
1 min readdistributed-systems · networking
2025
51 entriesFailure Detectors Cannot Be Perfect
In an asynchronous network you cannot distinguish a slow node from a dead one. Every detector picks a trade-off.
1 min readdistributed-systems · reliability
Read-Your-Writes Is a Product Requirement
Users notice when their own update disappears. Sticky sessions are a patch; version tokens are the fix.
1 min readconsistency · distributed-systems
Idempotency Keys Are a Contract, Not a Header
Retries are inevitable. A key plus stored response turns an unsafe request into a safe replay.
1 min readdistributed-systems · api-design
Exactly-Once Delivery Is a Transaction, Not a Transport
Queues promise it, but the guarantee always comes from deduplication and atomic state changes downstream.
1 min readdistributed-systems · messaging
Two-Phase Commit and Its Blocking Problem
2PC guarantees atomicity across databases and blocks forever if the coordinator dies at the wrong moment.
1 min readdistributed-systems · databases
Sagas Trade Atomicity for Compensation
A sequence of local transactions with compensating actions keeps availability and forces you to design for partial failure.
1 min readdistributed-systems · architecture
The Transactional Outbox Removes the Dual-Write Problem
Writing to a database and a broker in one request is not atomic. Write the event to the database instead.
1 min readdistributed-systems · messaging
Event Sourcing: Where Teams Get Stuck
The log is easy. Schema evolution, snapshots, and replay semantics are the hard parts.
1 min readarchitecture · databases
Vector Clocks Detect Concurrency, They Do Not Resolve It
A vector clock tells you whether two events are ordered or concurrent. The application still has to choose.
1 min readdistributed-systems · consistency
Lamport Timestamps and the Illusion of Order
A scalar logical clock preserves causality but cannot tell concurrency from causality.
1 min readdistributed-systems · consistency
Byzantine Fault Tolerance Has a Different Threat Model
Crash faults assume nodes stop. Byzantine faults assume they lie — and that changes the quorum math.
1 min readdistributed-systems · consensus · security
LSM Trees: Writes First, Reads Later
Buffer writes in memory, flush sorted files, and pay for it at read and compaction time.
1 min readstorage · databases
B-Trees vs LSM Trees
One updates in place, the other appends. The choice shapes write amplification, read latency, and space usage.
1 min readstorage · databases
Write Amplification Is the Hidden Tax on Storage
Every byte you write may be rewritten several times before it settles. SSDs charge you for all of it.
1 min readstorage · performance
The WAL Is Only as Durable as Your fsync
A write-ahead log makes crash recovery possible — if the record actually reaches stable storage before you acknowledge.
1 min readstorage · reliability · databases
When fsync Lies
Disks with volatile write caches, filesystems that ignore flushes, and virtualization layers that swallow them.
1 min readstorage · reliability
Copy-on-Write: Cheap Snapshots, Expensive Fragmentation
Never overwrite in place; write new blocks and flip pointers. Snapshots become free, and so do the problems.
1 min readstorage · filesystems
Erasure Coding Beats Replication on Storage, Not on Latency
k data shards plus m parity shards survive m losses for a fraction of the bytes — at the cost of reconstruction.
1 min readstorage · reliability
Compaction Strategies: Leveled vs Tiered
Leveled compaction keeps reads fast; tiered keeps writes cheap. Every LSM engine picks a point between.
1 min readstorage · databases
MVCC: Readers Never Block Writers
Keep multiple versions and hand each transaction a snapshot. The cost is version cleanup and write conflicts.
1 min readdatabases · consistency
Object Storage Consistency: What You Can Actually Rely On
Modern object stores are strongly consistent for new writes, but listings, caches, and multipart uploads have edges.
1 min readstorage · distributed-systems
Backpressure Is a Protocol, Not a Queue Size
When producers outrun consumers, the only choices are drop, buffer, or slow down. Pick deliberately.
1 min readnetworking · reliability
Head-of-Line Blocking, From TCP to HTTP/3
One lost packet can stall unrelated requests if they share an ordered stream. Multiplexing only helps if it is real.
1 min readnetworking · performance
Nagle Meets Delayed ACK: 40 ms You Did Not Budget
Two reasonable optimizations combine into latency spikes for small writes.
1 min readnetworking · performance
TCP Fast Open Saves a Round Trip and Rarely Gets Used
TFO carries data in the SYN, but middleboxes, cookies, and replay concerns limit deployment.
1 min readnetworking · performance
HTTP/2 Prioritization Failed, and That Is Instructive
A flexible priority tree was ignored by implementations; HTTP/3 replaced it with simple, explicit priorities.
1 min readnetworking · performance
QUIC Loss Recovery Without TCP Head-of-Line Blocking
Per-stream delivery means a lost packet costs one stream, not the connection — but congestion control still reacts globally.
1 min readnetworking · performance
Anycast Is Routing, Not Load Balancing
The same IP announced from many places gets users to the nearest one — and makes failures global.
1 min readnetworking · reliability
Connection Pools Are a Load-Balancing Decision
A pool per instance multiplies connections; a pool per cluster needs a proxy. Either way, size by concurrency, not by hope.
1 min readnetworking · databases · performance
Retries Turn a Blip Into an Outage
Synchronized retries amplify load exactly when the system is weakest. Jitter and budgets are not optional.
1 min readreliability · networking
Timeouts Are Not Latency Budgets
A per-call timeout without a deadline propagates work you no longer need. Deadlines are the fix.
1 min readreliability · networking
Tail Latency Is a Fan-Out Problem
If one request touches 100 services, the p99 of each becomes the p50 of the whole.
1 min readperformance · reliability
Little's Law: The Only Capacity Formula You Need First
Concurrency equals throughput times latency. Every queue-sizing argument follows from it.
1 min readperformance
Coordinated Omission Ruins Load Tests
If your load generator waits for a response before scheduling the next request, it hides the latency spikes users actually see.
1 min readperformance · testing
Never Average Percentiles
Percentiles are not additive. Aggregating p99s across shards or minutes produces a number that means nothing.
1 min readperformance · observability
Cache Stampedes and the Thundering Herd
When a hot key expires, every request misses at once. Single-flight and jittered TTLs fix it.
1 min readperformance · caching
The Thundering Herd Is a Synchronization Bug
Aligned timers, synchronized reconnects, and simultaneous retries all convert a small event into a spike.
1 min readreliability · performance
Circuit Breakers Are for Fast Failure, Not for Healing
They stop sending traffic to a failing dependency so it can recover, and so your own threads do not pile up.
1 min readreliability · architecture
Load Shedding: Serve Fewer Requests, Correctly
Under overload, dropping work early beats serving everything badly — if you drop the right work.
1 min readreliability · performance
Amdahl's Law and the Serial 5%
Parallelism helps until the serial fraction dominates. Most systems have more of it than the diagram suggests.
1 min readperformance
The Page Cache Is Your Best Database Cache
The kernel already keeps recently read files in RAM. Fighting it with application caches usually loses.
1 min readperformance · storage
JWT Pitfalls That Keep Coming Back
Algorithm confusion, unchecked claims, and treating a signed token as authorization.
1 min readsecurity · auth
PKCE Is Mandatory Now, Even for Confidential Clients
The authorization code flow is only safe if the code cannot be replayed by an interceptor.
1 min readsecurity · auth
Session Fixation: Rotate the ID at Login
If an attacker can plant a session identifier, the victim authenticates the attacker into their own account.
1 min readsecurity · auth
SSRF: When Your Server Fetches the Wrong URL
Any feature that fetches a user-supplied URL can be turned into a proxy into your private network.
1 min readsecurity · networking
CSRF in 2026: Mostly Solved, Still Misconfigured
SameSite cookies plus origin checks close the classic attack. Custom headers and token endpoints still need care.
1 min readsecurity · web
Supply Chain: You Are What You Install
Dependencies run with your privileges. Lockfiles, provenance, and least privilege are the controls that matter.
1 min readsecurity · reliability
Sandbox Escapes Start With a Shared Kernel
Containers isolate processes, not kernels. If the workload is hostile, that distinction is the whole threat model.
1 min readsecurity · infrastructure
Seccomp: Shrinking the Syscall Surface
Most workloads need a few dozen syscalls. Filters can make the rest fail fast.
1 min readsecurity · infrastructure
Capabilities Beat Root-vs-Not-Root
Linux splits root privileges into capabilities. Grant the few you need and drop the rest.
1 min readsecurity · infrastructure
Zero Trust Is an Architecture, Not a Product
Authenticate and authorize every request based on identity and context — and assume the network is hostile.
1 min readsecurity · architecture
2024
10 entriesZero-Knowledge Proofs: Prove Without Revealing
A proof of a statement reveals nothing but its truth. The engineering question is what the statement costs to prove.
1 min readcryptography · zero-knowledge
zk-SNARKs vs zk-STARKs
Trusted setup and small proofs versus transparency and larger proofs. The choice is about assumptions, not speed alone.
1 min readcryptography · zero-knowledge
ML-KEM: The Post-Quantum Key Exchange Standard
Kyber became ML-KEM. Here is what changes in a TLS handshake and why hybrid mode is the migration path.
1 min readcryptography · post-quantum · tls
ML-DSA and the Signature Migration
Dilithium is now ML-DSA. Larger signatures, different performance, and a hybrid period ahead.
1 min readcryptography · post-quantum
Homomorphic Encryption: Compute on Ciphertext
FHE is real and still expensive. Know which variant matches your operation before committing.
1 min readcryptography · privacy
Secure Multi-Party Computation Without the Hype
Multiple parties compute a joint function without revealing inputs. The costs are rounds and bandwidth.
1 min readcryptography · privacy
Verifiable Delay Functions Force Sequential Time
A VDF takes a known number of sequential steps to compute and is fast to verify — a clock made of math.
1 min readcryptography · distributed-systems
Threshold Signatures Remove Single Points of Key Failure
t-of-n parties produce one signature without ever assembling the full private key.
1 min readcryptography · security
Ring Signatures: Sign on Behalf of a Set
A signature proves a member of a group signed, without revealing which one. Anonymity with a fixed cost.
1 min readcryptography · privacy
Verkle Trees Shrink Witnesses
Vector commitments replace per-level hashes, so proofs stay small as the tree widens.
1 min readcryptography · storage