Existing contracts. Different machinery.
The useful conversion test preserves what the outside world expects, changes the distributed implementation underneath it, and reports the behavior and cost separately.
Redis-on-Ribbit.
Redis’s own external behavior is the oracle. Published checkpoints include 1,502 OK / 207 N/A on one core and a separate ARM64 parallel run with unresolved errors. Internal bookkeeping or synchronous reader handoff that the new model intentionally declines must be identified explicitly rather than counted as a pass.
The original site’s Pi 5 throughput record reports Ribbit SET 268k/s on one core and 704k/s on three; GET 399k/s and 858k/s; LPUSH 424k/s and 1.03M/s. Those are Ribbit rates for that million-key workload, not pairwise speed ratios. A small-cache i7 measured slower than stock Redis in several controls.
Scaling by changing state instead of moving ownership.
| Test | Stock Redis | Ribbit front | Interpretation |
|---|---|---|---|
| RT7: 100k keys, 3→4 primaries | 24,936 keys moved; 23.3 s | No keys moved; 0.067 s | A capacity/topology-change experiment, not per-command latency. |
| RT3: replica join, delay 0, 10k keys | 0.019 s | 0.057 s | Stock was faster at this size. |
| RT3: same, 1M keys | 2.1 s; 23.9 MB copy | 0.059 s; about 2 KB | The tested front changes shared configuration instead of copying the key population. |
Reader errors were zero in both RT7 arms. These values are preserved as reported experimental records; the website update has not repeated those workloads.
lispers.net.
The conversion preserves the wire contract while moving registration and resolution truth into owned state. It includes multi-record Map-Register with all-or-nothing behavior, registration refresh, prefix lookup and freshness-based ETR liveness. The v0.44 record has a 42/42 conformance gate; earlier broader acceptance campaigns used different counts and are not summed with it.
The published code measurement compared 26 original lisp files at 22,527 LOC / CCN 6,764 with four conversion files at 2,274 LOC / CCN 1,215. Shared infrastructure was measured separately: 14 files, 3,477 LOC / CCN 2,046. Keep that denominator visible when describing the reduction.
RINA / rlite.
The shared-state experiment moves credit and route information into the region while preserving reliable flow behavior. The recorded ring/window suite passed 30/30. In a triangle test with an intermediate member killed, the DIF healed in about 40–43 ms; the reliable flow lost 0/100 PDUs, and the unreliable flow lost one in flight.
A separate relay run increased throughput from about 635 to 1,243 Mbps with 20,000/20,000 delivered after the credit-state change. Those are measurements of that run. FD&F’s newer namespace suites exercise a different hybrid and transport arrangement.
PyTorch.
PyTorch above the distributed backend remained PyTorch. Conservative shared-state arms preserve ordinary DDP agreement; fresh-only and every-k experiments deliberately change consumption and must be kept separate. Read the campaign data and controls.
What the conversions taught.
Address ownership, held views and publication marks are recurring tools across different domains. The implementations become smaller because work and choreography have been removed. That is a reason to investigate another system, not a guarantee that an arbitrary rewrite will be smaller or faster.
The supplied JavaScript programming guide carries the same rules into a new binding. Its examples include latest/distinct state, held queries, claims, a foreign source and restart. Read the rules before making the conversion.