Benchmarks¶

Bindizr measured against PowerDNS Authoritative, Technitium DNS, Knot DNS, CoreDNS, and plain BIND9 (nsupdate / rndc) on identical hardware, datasets, and container limits — the suite lives in benchmarks/. Every figure is the mean of 5 runs on an 8-core AMD Ryzen 7 9800X3D, each container capped at 4 CPU / 4 GB.

No overhead on the query path¶

DNS query throughput: CoreDNS 62,696 QPS, Bindizr + BIND9 62,448, Native BIND9 61,629, PowerDNS 60,406, Knot DNS 41,289, Technitium 13,139

Bindizr never answers a client query — the BIND9 secondaries do. Bindizr + BIND9 serves 62,448 QPS against native BIND9's 61,629 (−1.3%, within run-to-run noise), and Bindizr itself draws 0.7% CPU under that load.

Bulk import¶

Bulk import of 10,000 records: Bindizr zone file 132,720 records/sec, BIND9 + rndc 102,641, Bindizr bulk API 93,364, PowerDNS 36,443, Knot DNS 18,514, CoreDNS 10,374, Technitium 9,244

A 10,000-record zone file imports in 76 ms; the same records through the bulk record API take 107 ms. Both paths commit to the database, so the zone survives a restart and transfers to the secondaries immediately.

Incremental transfers stay incremental¶

IXFR transfer size in a 100,000-record zone: Bindizr moves 736 B for 1 change up to 558 KB for 10,000 changes, while PowerDNS moves about 5.5 MB regardless of the change count

A snapshot per SOA serial means an IXFR carries only what changed: 736 B for a single change in a 100,000-record zone, where the full zone is 5.5 MB. PowerDNS answers the same request with the entire zone; Knot DNS and Technitium track the Bindizr curve.

Write path¶

Median record-create latency from API call to DNS visibility: Technitium 0.6 to 4.6 ms, PowerDNS 2.8 to 7.3 ms, Knot DNS 16.2 to 20.4 ms, BIND9 + nsupdate 17.2 to 21.5 ms, Bindizr + BIND9 6.4 to 65.7 ms

A create is acknowledged in 6.4 ms and answers from the secondaries 65.7 ms after the call (p95 83 ms, no timeouts). Bindizr commits to the database and propagates by NOTIFY + IXFR, where the integrated servers answer from their own process as soon as they accept the write.

Record CRUD throughput¶

System	Create TPS	Update TPS	Delete TPS	Read TPS	Read p95	Errors
Bindizr + BIND9	198.1	197.1	186.8	13,916.9	2.82 ms	0.00%
Technitium DNS	8,992.2	8,182.1	9,123.3	10,594.2	3.51 ms	0.00%
Knot DNS	759.2	1,116.3	735.1	1,203.1	36.73 ms	0.00%
BIND9 + nsupdate	412.8	948.2	411.4	1,199.7	36.92 ms	0.00%
PowerDNS Authoritative	87.4	72.1	83.9	2,962.0	4.27 ms	0.00%

Each write is a durable database commit plus a zone-serial bump, which sets the per-record write rate — servers that hold the zone in memory do more per second here. These runs use SQLite; PostgreSQL raises creates from 199 to 571/sec. Read is a management-plane read: an API GET where there is an API, a dig subprocess otherwise, so those p95s carry process-spawn cost.

Database backends¶

Backend	Create TPS	Read TPS	Read p95	100k bulk import	Peak memory
SQLite	198.6	14,236.1	2.78 ms	0.96 s (104,391/sec)	120 MB
MySQL	262.0	12,766.3	4.35 ms	2.04 s (48,994/sec)	879 MB
PostgreSQL	571.3	12,432.4	4.23 ms	1.98 s (50,647/sec)	325 MB

Bulk import stays near-linear from 10k to 100k records on all three backends.

Software under test

Software	Version
BIND9	`internetsystemsconsortium/bind9:9.21`
Bindizr	built from source
CoreDNS	`coredns/coredns:1.14.6`
Knot DNS	`cznic/knot:3.5`
MySQL	`mysql:9.7`
PostgreSQL	`postgres:18`
PowerDNS	`powerdns/pdns-auth-49`
Technitium	`technitium/dns-server`