Best-Effort Both Ways

The catalogue's two pure seats are mirror images, and reading them against each other is the fastest way to feel the trade. Best-effort availability guarantees consistency and optimizes responsiveness: Chubby, Google's lock service, runs a Paxos-replicated database behind a small-file face, one cell per datacenter, read-optimized around a stable primary — strongly consistent always, and unavailable only when the cell splits. Best-effort consistency guarantees the answer and bounds the freshness: the web cache Akamai pioneered places content in datacenters worldwide and serves each reader from a nearby one, so an answer is nearly always fast, and a page that is slightly out of date beats a page that never loads. One guarantees correctness and merely optimizes speed; the other guarantees speed and merely optimizes correctness.

That is the whole lesson of the pair: you guarantee one property and optimize the other — never both — and which one you guarantee is set less by taste than by where the system stands. Chubby's seat is right inside one datacenter, where links are fast and reliable and a primary failover is rare, so guaranteeing consistency costs almost no availability in the common case. Akamai's seat is right across the world, where users are dispersed and a strongly consistent answer would degrade rapidly, so guaranteeing availability forces consistency down to best-effort. This platform seats both at once — the ledger and the gates take Chubby's seat at the consistent core, the memory, bus, and surface take Akamai's at the available edge — and the branded seam of E2 is what lets one system hold two opposite seats and stay one system.

The two systems are not opposites of temperament; they are opposites of place. Chubby is deployed one cell per datacenter, where communication is fast and reliable and the failure of a primary is infrequent. In that setting, guaranteeing consistency costs almost nothing in availability — the network is kind nearly always, so a system that blocks under a split still answers promptly the rest of the time. Its engineering leans all the way in: read-only requests are prioritized and served without the replicated log incurring extra round trips, so correctness and speed coexist in the common case. Consistency is the guarantee because, right there, it is affordable.

Akamai lives at the other pole — content delivered to users scattered across the planet, where a strongly consistent read would have to reach back to an origin and pay wide-area latency, and where mobile links make that worse. There, availability is the only thing worth guaranteeing: serve from the nearest cache and let freshness be best-effort, because a fast, slightly stale answer is precisely what the reader wants. The platform mirrors this in its own floor plan: the ledger and gates sit at a coordinated core and take the consistency seat; the memory, bus, and surface sit near the reader and take the availability seat. Geography picks the seat at each tier, and the identity threaded through all of them is what keeps the arrangement a single system rather than two.

The two-seats rule — you guarantee one property and optimize the other, never both: Chubby guarantees correctness and optimizes availability, the web cache guarantees an answer and optimizes freshness. Which seat a tier takes is set by where it stands — consistency is cheap to guarantee at a coordinated core, availability is cheap to guarantee near the reader — and a system holds both only because one identity crosses between them.