Performance,
in full.

Timing columns show µs/op. The p95 is a percentile of sample averages, not individual-event tail latency. Highlights compare participants in the same scenario and metric.

create/10

Create an editable 10-item graph; cleanup is outside timing.

create+encode/10

Create an editable 10-item graph and export a transferable initial state; cleanup is outside timing.

create/1000

Create an editable 1000-item graph; cleanup is outside timing.

create+encode/1000

Create an editable 1000-item graph and export a transferable initial state; cleanup is outside timing.

export-first/1000

Export a prepared 1000-item document for the first time; construction, import verification and cleanup are outside timing.

export-repeat/1000

Export a prepared 1000-item document after one untimed export, without any edit; construction, import verification and cleanup are outside timing.

read-count/10

Read a cached editable nested count; initial container lookup is outside timing.

mutate/history-0

One committed count update through each library's default local API; Pulsify retains its native patch without encoding. Encoding is measured separately in edit+encode.

mutate/history-1000

One committed count update through each library's default local API; Pulsify retains its native patch without encoding. Encoding is measured separately in edit+encode.

mutate/history-5000

One committed count update through each library's default local API; Pulsify retains its native patch without encoding. Encoding is measured separately in edit+encode.

list-append/10

Append a new editable two-field object to the list and commit once per append (adapted from B1.8; objects replace upstream's numbers).

edit+encode/10

Capture a version, commit one count edit, encode its incremental update; verify all encoded packets on a peer outside timing.

delta-roundtrip/10

Version-vector capture + edit + delta encoding + peer import + peer count read; synchronous CPU-only, without transport or acknowledgement.

delta-roundtrip/1000

Version-vector capture + edit + delta encoding + peer import + peer count read; synchronous CPU-only, without transport or acknowledgement.

initial-load+materialize/10

Construct an empty replica, import an initial state and materialize a detached JSON view. Export is prepared outside timing; cleanup is outside timing.

initial-load+materialize/1000

Construct an empty replica, import an initial state and materialize a detached JSON view. Export is prepared outside timing; cleanup is outside timing.

offline-merge/100-per-peer

Two initialized peers make 100 committed offline edits each; time vector capture, two-way delta encode/import and materialization. Edits are outside timing.

upstream-map-conflict/16-peers

Adapted B3.1: 16 peers sharing a baseline concurrently update one map key; time importing all deltas and materializing one peer. Setup/encoding and a reference merge are outside timing.

proxy-overhead/read-count

Same cached nested field, native API versus the Pulsify pipeline bypassing proxies versus the public proxy.

proxy-overhead/mutate-history-0

Native json-joy versus identical Pulsify commit/native-journal pipeline with and without proxy traps; no encoding during local writes, encoded rows are verified identical after export.

proxy-overhead/mutate-history-1000

Native json-joy versus identical Pulsify commit/native-journal pipeline with and without proxy traps; no encoding during local writes, encoded rows are verified identical after export.

proxy-overhead/mutate-history-5000

Native json-joy versus identical Pulsify commit/native-journal pipeline with and without proxy traps; no encoding during local writes, encoded rows are verified identical after export.

Timing columns show µs/op. The p95 is a percentile of sample averages, not individual-event tail latency. Highlights compare participants in the same scenario and metric.

Each operation’s rate is its completed count divided by the whole mixed batch duration, not standalone operation capacity.

The primary large native comparison uses three baseline and three optimized campaigns with identical fixtures. Docker and medium each use one before/after pair, with less evidence about variance. These are separate from the original engine comparison.

Historical Docker / Bun suite: 3 warmups + 12 measured samples per case. FoundationDB used a Python JSON/base64 pipe per primitive; its transport costs are included. These scores are separate from the native Rust campaign above.

Bun 1.4.4 / PostgreSQL 16.15 over localhost TCP in Docker. Durable logged writes, 128-byte opaque payloads, 2 warmups and 6 measured runs. The statistic is the fourth sorted sample of six (upper middle), rather than the arithmetic median. Batch sizes are different workloads and are not ranked against one another.