KV Usage

Key-Value store provides low-latency global data: journeys, cached recommendations, intermediate article sets, profile photo prompts.

Consistency Model

Eventual consistency; design assumes tolerant reads. Critical persistence (articles/prompts) sent to Mantle2 instead.

Patterns

• Read-through: tryCache(key, fetcher)
• Write-back: compute -> store -> respond
• Expiration via TTL embed (stored metadata)

Serialization

Compact JSON objects; avoid deeply nested arrays. Example:

{
	"recommendations": ["000000000000001234567890", "000000000000001234567891"],
	"generated_at": 1731206400
}

All IDs use 24-digit zero-padded format.

Size Constraints

• Keep values < 100KB for latency
• Split large arrays into shards (<25KB each)

Failure Handling

• Soft fail on KV read (fallback to source fetch)
• Log latency >200ms for tuning