hard_drive

Storage
Primitives

Data has gravity. Moving it is expensive. Understanding the physical medium—spinning rust vs. flash chips—determines your database choices.

The Storage Hierarchy

Registers

< 1nsBytes

L1/L2 Cache

~10nsKB/MB

Main Memory (RAM)

~100nsGB

SSD (Flash)

~100µsTB

HDD (Magnetic)

~10msPB

"Faster storage is smaller and more expensive. Slower storage is massive and cheap. Caching is just moving data up this pyramid."

Sequential vs Random Access

fast_forward

Sequential

Reading blocks of memory one after another. No seeking.

✓ Predictable pre-fetching
✓ Extremely fast on both HDD & SSD
✓ Used by Kafka logs, LSM Trees

shuffle

Random

Jumping to arbitrary locations on disk. Requires seeking.

⚠ Slow on HDD (physical arm movement)
~ Okay on SSD (but still slower)
⚠ Used by B-Trees, Hash Indexes

Rule of Thumb: Sequential access on HDD is faster than Random access on SSD.

HDD vs SSD Mechanics

HDD (Hard Disk Drive)

Spinning magnetic platters with a mechanical arm.

Seek Time: Moving the arm (latency). ~10ms.
Rotational Latency: Waiting for the sector to spin under the head.
Best for: Archival data, Sequential logs.

SSD (Solid State Drive)

NAND Flash memory. No moving parts.

IOPS: Can handle 100x more random IOPS than HDD.
Wear Leveling: Flash cells degrade after writes.
Best for: Databases, Boot drives, Application servers.

Durability & WAL

save

Write-Ahead Logging (WAL)

Writing to disk is slow. Writing to memory is volatile. How do databases survive a crash without being slow?

1. Append to LogSequential write to disk. Fast. (The WAL)

arrow_forward

2. Update MemoryModify the in-memory structure (B-Tree/Memtable).

arrow_forward

3. Async FlushLater, write memory to disk data files.

If crash occurs, replay the WAL to reconstruct memory state.

B-Trees vs LSM Trees

B-Trees

Standard in Postgres, MySQL.

Opt: Read heavy workloads.
Cons: Disk seek for every leaf. Random writes.
"Updates in-place. Highly consistent but slow for massive writes."

LSM Trees

Cassandra, DynamoDB, LevelDB.

Opt: Write heavy workloads.
Cons: Compaction overhead. Slower reads.
"Append-only. Leverages sequential writes for speed."

Architect's Summary: Use B-Trees for stability and read speed (Postgres). Use LSM Trees for extreme write volume (Cassandra).