Choosing a database: SQL, documents, graphs, and search
Key-Value Stores
The fastest database lookup possible β by exact key
A key-value store works like a massive dictionary lookup. You walk up, hand over a key, and instantly get the value back. No scanning rows, no joining tables, no evaluating conditions. You know the key β you get the value in under a millisecond. Nothing is faster for this pattern.
Engineers discovered that most reads in large-scale systems are simple lookups: fetch the user session, get the feature flag, check the rate limit counter, retrieve the cached API response. None of these need SQL. They just need a key. Inventing a dedicated engine for this pattern made it 100x faster and 10x cheaper to run.
Redis
The ultra-fast in-memory data structure server
Redis
The ultra-fast in-memory data structure server
Mental model
Redis acts like ultra-fast temporary memory for your application. Imagine your server spending 200ms on a complex database query that gets executed 1,000 times per second. With Redis, you calculate it once, store the result in memory, and serve all 1,000 requests from Redis in under 1ms each. The expensive work happens once; Redis handles the rest.
Why it exists
Created in 2009 by a Sicilian developer who was building a real-time web analytics tool and found MySQL couldn't handle the write throughput. Redis was designed to hold all data in RAM, making reads and writes orders of magnitude faster than disk-based databases.
Shines at
- Session storage and authentication tokens
- Caching expensive query results
- Real-time leaderboards and sorted sets
- Rate limiting and request throttling
- Pub/sub messaging between services
- Job queues and background task management
Breaks down when
- Data larger than available RAM (expensive at scale)
- Complex query patterns beyond key lookups
- Long-term durable primary storage
- Relational data with joins
Real-world use
Twitter uses Redis for its timeline cache. GitHub uses Redis for rate limiting. Stack Overflow caches nearly everything through Redis.
DynamoDB
Amazon's infinitely scalable managed key-value store
DynamoDB
Amazon's infinitely scalable managed key-value store
Mental model
DynamoDB is designed around one promise: regardless of how many users you have, a DynamoDB read will respond in single-digit milliseconds. It achieves this by distributing your data automatically across many servers based on a partition key β so no single server ever becomes a bottleneck.
Why it exists
Amazon built DynamoDB after their 2004 internal crisis: as Amazon.com grew, their Oracle database couldn't keep up. They needed a system that scaled essentially infinitely, never went down, and required zero database administration. DynamoDB was born from those requirements.
Shines at
- Serverless and AWS-native architectures
- Gaming leaderboards and user profiles at massive scale
- Shopping carts and session state
- IoT device state management
- Workloads with predictable access patterns by partition key
Breaks down when
- Complex queries beyond key-based lookups (expensive with indexes)
- Ad-hoc analytical queries
- Relational data models with joins
- Cost at very high read/write throughput without optimization
Real-world use
Amazon.com uses DynamoDB for shopping carts. Lyft uses it for geolocation. Snapchat uses it for messaging metadata at 10B+ messages per day.
Your Path
Key Traits
Pick When
- Cache in front of slow DB
- Session / auth token store
- Leaderboard / rate limiter
Used By
Core Tradeoff
Extreme speed β data must fit keyβvalue mental model
Browse database families
Decision explanations
What operational model best fits this workload?
A platform records thousands of financial operations per minute. Each operation touches multiple tables β deducting from one account, crediting another, updating a balance ledger. Partial writes must never succeed. Queries need to join across user, account, and transaction tables in real time.
Transactional (OLTP)
OLTP = individual, frequent writes where every single transaction must be correct. OLAP is the opposite β it scans huge amounts of historical data to answer questions like 'what was total revenue last quarter?' When every write matters, you need OLTP.
Analytical (OLAP): OLAP databases are built for scanning large amounts of old data all at once (like monthly reports). They trade away per-row correctness for bulk scan speed β terrible for thousands of individual write operations per second.
What storage architecture optimises this workload?
You run nightly BI reports: total revenue by region, cohort retention, anomaly detection. Queries scan hundreds of millions of rows but nothing is ever updated after it's written. Tables have 80+ columns per row but queries only need a few at a time.
Data Warehouse (Columnar)
Columnar storage is a direct response to analytical query patterns. When you query 2 out of 80 columns across 500M rows, a row-store reads all 80 columns for every row β wasting 97% of disk I/O. A columnar store reads only those 2 columns, compresses them heavily, and aggregates them in a single scan pass.
RDBMS (Row Store): Row-oriented stores are fast at fetching complete records β but analytical queries scan billions of partial rows. An RDBMS touching 500M rows would deserialise all 80 columns per row even when you only need 2 β wasting enormous I/O bandwidth and making queries 10β100Γ slower than a columnar store.
Which transactional database family fits this workload?
Your app handles 8,000 transactions per second in a single region. Schema is stable. Queries join across 5β8 tables per request. No need to spread writes across multiple servers.
RDBMS
NewSQLβs distributed write scaling adds operational complexity only justified when you need to shard writes across regions β typically at 100k+ TPS or for global multi-master deployments. This workload fits well within primary-replica RDBMS capacity, making NewSQL overhead unnecessary.
NewSQL: NewSQL solves one problem: writing to multiple servers across regions simultaneously. Great at Google scale. But it adds latency overhead per write and is complex to operate. If one server handles your load, RDBMS is the simpler, faster choice.
What data organisation model fits a workload where every entity carries a different field set?
Your product catalog has 200M items. Electronics have wattage. Clothes have size and material. Furniture has dimensions. No two categories share the same fields. 40% of queries fetch a full item. 40% filter by attribute combinations. Schema changes every time a new category launches.
Document-Oriented
Document databases let each item carry exactly its own fields β no null columns for attributes a product doesn't have. You can add an index on any field and query it directly. Key-value stores can only look up by exact key β you'd need to build an entire query engine on top just to filter by attributes.
Key-Driven Access: Key-value = 'give me the item with key #12345'. That's the only query it supports. No 'show me all electronics under $100', no filtering by attributes. For a query-heavy product catalog you'd have to pre-compute every possible filter combination yourself β that's the application's job, not the database.
As multi-hop relationship traversal becomes the dominant query pattern, what architectural response is correct?
Your social app runs on MongoDB. A new feature needs: 'find friends-of-friends who bought the same product the user did' β that's 4 levels of relationships. These queries take 9β14 seconds and pin the CPU. The deeper the social graph grows, the worse it gets.
Graph Database
Graph databases store relationships as physical pointers. Each hop is O(1) β instant, no matter how big the database is. In MongoDB, every hop is another aggregation step that multiplies the work. At 4 hops on 10M users, that's billions of operations. Graph databases solve this in milliseconds.
Denormalised Documents: Embedding relationships in arrays works for simple things like 'a post's top 5 comments'. Social graphs are different β they're unbounded. A user's 3-hop network could be millions of people. Storing that as nested arrays uses gigabytes per user and the query is still slow. It doesn't fix the fundamental problem.
For this read-dominated, key-exact workload, which architecture serves the latency and access pattern?
A ride-hailing app tracks live GPS for 3M drivers (200k writes/sec, 1M reads/sec, must respond in under 2ms). Also serves surge pricing per city zone and manages session data per active ride. All lookups are exact key lookups β no filtering or joins needed.
In-Memory Key-Value Store
The bottleneck is read speed: under 2ms. In-memory key-value stores (like Redis) keep all data in RAM β reads are microseconds, not milliseconds. Column-family databases write to disk, so reads take 15β50ms. When you need blazing-fast reads, you need data in memory.
Distributed Column-Family: Cassandra excels at absorbing massive write storms β IoT sensors writing millions of events per second. The trade-off: reads hit disk, taking 15β40ms. Perfect for 'store everything and query later' data. Terrible for a live driver map that needs to respond in under 2ms.
When time is the organising dimension and every query is a time range, what database family is architecturally aligned?
80,000 connected vehicles each send 400 sensor readings per second (speed, battery, brakes, GPS...) = 32M data points per second. You need: 'plot battery over the last 4 hours for vehicle X', 'alert if brake pressure drops', 'fleet average fuel consumption over 7 days'. No text search. Every query is a time window.
Time-Series Database
Time-series databases are built exactly for sequential numeric measurements. They compress repeated similar values efficiently (battery at 80%, 81%, 80%... is stored as tiny deltas) and run time-window queries natively. Search engines index words β vehicle sensor data has no words. You'd pay enormous memory cost for indexes that are completely useless here.
Search / Log Analytics Engine: Search engines shine for text: log messages, error descriptions, user queries. They build word indexes. But vehicle telemetry is pure numbers β speed: 62, battery: 77%, brake: 94%. There's nothing to text-search. You'd waste memory on word indexes and miss the compression and time-range query features that time-series gives you.
