MariaDB restore with Resistro CLI v1.6.72: 1.9x faster than the mariadb client
Same 380 MiB archive, two ways back: measured wall clock, memory use and phase log, verified by row counts and table checksums.
Our earlier post about the MariaDB restore described two separate phases: first parse the SQL archive into TSV serially, then load it in parallel. That path no longer exists. Since Resistro CLI v1.6.70 the streaming restore is the default: the parser reads the archive and hands finished chunks to the loaders while they are still loading. That also makes the old figure obsolete, because it compared two versions of the two-phase path with each other.
This measurement therefore compares against the path people otherwise take: the mariadb client replaying the archive.
Note (2026-11-19): this post replaces the earlier measurement for v1.6.48.
Test setup
| Item | Value |
|---|---|
| Archive | mariadb-dump plain text, 398,882,644 bytes (380.4 MiB), 13 tables, largest 2,000,000 rows, about 3.02M rows in total |
| Server | MariaDB 11.8.9 in a container, data directory in RAM (tmpfs), default configuration |
| Host | 8 cores, 15 GB RAM; the host disk is not involved in the restore |
| Resistro | CLI v1.6.72 (commit dce3a75), restore single <archive> --target <new> --create --confirm --jobs 8 --no-analyze --skip-disk-check |
| Reference | mariadb -uroot <archive> in the same container, same archive |
| Runs | 3 per variant, interleaved (A, B, B, A, A, B), fresh container state before each run |
| Timing | /usr/bin/time -f "%e %M", wall clock and maximum RSS of the process |
Result
| Variant | Run 1 | Run 2 | Run 3 | Median |
|---|---|---|---|---|
mariadb client |
24.82 s | 26.15 s | 25.20 s | 25.20 s |
| Resistro v1.6.72 | 14.48 s | 13.26 s | 13.41 s | 13.41 s |
25.20 / 13.41 = 1.88. The restore takes 46.8 % less wall clock with Resistro than with the client. In medians that is 28.4 MiB/s versus 15.1 MiB/s.
Peak RSS was a constant 27.4 MB for the client and between 97.9 and 101.8 MB for Resistro. That is the price of the parallel loaders, each holding its own buffers.
What the phase log shows
After each restore Resistro prints one line with the phase timings. All three runs of this setup:
restore_phases parse_ms=0 split_ms=0 connect_ms=8 schema_ms=17 load_ms=12401 post_ms=1 sum_ms=12429 total_ms=12429
restore_phases parse_ms=0 split_ms=0 connect_ms=8 schema_ms=15 load_ms=12565 post_ms=1 sum_ms=12590 total_ms=12590
restore_phases parse_ms=0 split_ms=0 connect_ms=9 schema_ms=16 load_ms=13638 post_ms=1 sum_ms=13665 total_ms=13665
parse_ms=0 is not a measurement error, it is the architecture: there is no second pass over the full archive that parses ahead of the load any more. Nearly all of the time sits in load_ms.
Verification
After every restore, row counts and CHECKSUM TABLE of all 13 tables were compared against the source. For both variants the difference is empty — the faster path writes the same data.
What this does not say
- MariaDB only, this archive only. MySQL, PostgreSQL and other key and column types are not measured here.
- The data directory was in RAM. On a server with real disks the share spent waiting on IO shifts, and so does the gap.
- No comparison against
myloaderor other parallel third-party tools. The client is the reference because it is available everywhere without extra software. - A single host (8 cores, 15 GB RAM), three runs per variant, no repetition on other hardware.