SpaceDB Is 3.8 Percent Leaner in Its 0.7.2 Second Pass
SpaceDB 0.7.2 is a second instruction-count pass on the local-first database: 3.8 percent fewer instructions across fifteen drivers, with bigger cuts on scans and vectors.

What did SpaceDB 0.7.2 change compared with 0.7.1?
SpaceDB 0.7.2 is a second instruction-count pass. Fifteen repository drivers fell from 2,702.8 million to 2,599.0 million instructions against a re-measured 0.7.1, a 3.8 percent drop. Range scans, the console, vector search and query setup moved most. No wall-clock figure was reported.
SpaceDB 0.7.2 is the second, smaller performance pass on the local-first database, and the honest headline is a small number. Across fifteen measured drivers, total instructions fell from 2,702.8 million to 2,599.0 million against a freshly re-measured 0.7.1. That is 3.8 percent. A few specific jobs moved a lot more: range scans inside a write fell 36.1 percent, the operator console 33.1 percent, vector search 29.8 percent, and query setup 18.0 percent.
If you read our earlier SpaceDB 0.7 performance release, this is not that story again. That release was the big one, measured against 0.6.0. This one is a follow-up pass with a different baseline, a wider driver set, and much smaller totals. The crates are spacedb-sdk and spacedb-store at version 0.7.2, published on 7 November 2026 from public commit 43108abe.
If SpaceDB is new to you, start with SpaceDB as a local-first Firebase alternative. It is a CRDT-native database where the replica is your device. The product page is SpaceDB on MATA, and the builder notes are at remadewithrust.com.
Every figure below is repository-reported. We read the README and the commit. We ran nothing.
What The Second Pass Counted
The yardstick is the same kind as before: callgrind instruction counts from the Valgrind tool suite on deterministic drivers, with random seeds pinned so a count repeats. The README's reproduce line builds an example in release mode and runs it under callgrind on Linux. It does not name the CPU or the rustc version, and no wall-clock time is reported for this release.
The moves, largest first:
- Range scans inside a write transaction: 11.6 million to 7.4 million instructions, down 36.1 percent.
- Operator console rollups: 68.2 million to 45.6 million, down 33.1 percent.
- Vector search, 1,500 vectors of 64 dimensions, three metrics, top 10: 33.1 million to 23.3 million, down 29.8 percent.
- A WASM map-reduce query plus a host call: 33.4 million to 27.4 million, down 18.0 percent.
- One transaction per CRDT mutation: down 12.7 percent. The CRDT driver: down 10.6 percent.
- Access control: down 0.3 percent. Durability: no change. Store: down 1.5 percent.
These are instruction counts on fixed drivers, not end-to-end query latency or a timed speedup.
One footnote matters. The store driver starts at 194.1 million here, lower than where 0.7.0 ended, because 0.7.1 halved zstd decode cost. That drop belongs to 0.7.1. It is not a win for this release, and we do not count it as one.
Where SpaceDB 0.7.2 Spent Fewer Instructions
The changes are about not copying and not repeating. The database rules, the CRDT merge, and the consistency tiers are the same.
Range Scans Now Lend The Row
Now a range scan can lend each row to the caller through range_with, so the row does not have to be copied out unless the caller keeps it. Presence checks no longer copy the row either, and collections accept borrowed keys. The range-scan driver is the one that moved 36.1 percent.
Mutations Batch Into One Transaction
CRDT mutations now skip a cleanup observer that cloned three structures, and a batch of mutations can run in one transaction through CrdtDoc::batch. The storage layer opens each table once per transaction. A query hashes its snapshot and module once and builds its function linker once. Short internal keys use a cheaper non-cryptographic hash, and the console interns names.
Eight Lanes For Vector Math
Vector dot products and distances now sum in eight independent lanes, so the compiler can vectorize them, and then fold in a fixed order. The code is in the vector index. Floating-point addition is not perfectly associative, so a score can differ from 0.7.1 in its last bits. The project's test corpus ranked identically. Two candidates that sit within that rounding could swap places, and the README says so.
What The Pass Means For Data You Keep
Local data is the point of SpaceDB. A home machine that keeps the durable copy pays for every instruction itself, so cheaper work is a practical reason to keep data home. The connection to the three primitives of Digital Freedom, Trust, Security, Incentive, is direct and modest.
Trust: Read The Limits First
No wall clock, no named CPU, no named compiler. The public commit's subject names a monorepo SHA that is not a commit in the public repository, and the detailed per-copy inventory lives in that private monorepo. Wins measured inside the yrs, primeorder and rusty_zstd dependencies are explicitly left out of these numbers. Those limits are written down by the project, which is what makes the numbers usable.
Security: Nothing New To Claim
One stored-format detail changed. New rows skip zstd's optional content checksum, a field defined in the Zstandard format, RFC 8878, because the README says the AEAD tag already authenticates every row. Authenticated encryption is specified in RFC 5116. We did not audit SpaceDB's use of it for this article, so we claim no new guarantee. Older rows with the checksum still read.
Incentive: Cheaper Work On The Home Computer
The home computer era asks a household machine to hold real data. Every instruction it does not spend is time and energy it keeps. The NIST Privacy Framework is a reference for managing privacy risk that comes from processing data. Keeping that processing on a machine you own is one way to shrink the risk, and SpaceDB 0.7.2 makes the local side slightly cheaper. The Freedom Guide is the place to start, and Learn has the rest of the series.
FAQ
Short answers already in this article.
Is SpaceDB 0.7.2 the same story as the SpaceDB 0.7 release?
No. The large 0.7.0 drop is the earlier SpaceDB 0.7 article. SpaceDB 0.7.2 is measured against a re-measured 0.7.1 on fifteen drivers, and the halved zstd decode cost belongs to 0.7.1, not to 0.7.2.
Can SpaceDB 0.7.2 change vector search results?
Slightly, at the edges. Vector scores are now summed in eight lanes, so an f32 score can differ in its last bits from 0.7.1, and two near neighbors within that rounding could swap places. Rankings in the project's test corpus did not change.
Does SpaceDB 0.7.2 add a new security guarantee?
No new guarantee is claimed here. New rows skip zstd's content checksum because the README says the AEAD tag already authenticates them, and 0.7.2 still reads rows that carry the checksum. We did not independently audit that reasoning.
Which SpaceDB 0.7.2 drivers barely moved?
Access control moved 0.3 percent, durability did not move, the store driver moved 1.5 percent and the 40-replica simulator moved 1.8 percent. The big moves were range scans, the operator console, vector search and query setup.
Which crates carry the SpaceDB 0.7.2 pass?
spacedb-sdk 0.7.2 and spacedb-store 0.7.2, published on crates.io on 7 November 2026 from public commit 43108abe. The project reports cargo semver-checks passing against 0.7.1.

