Changed
- Agent: a sampler for a capability the machine lacks now reports
unsupported
rather thanfailed, andrezolus statusno longer exits non-zero because of
it. A host whose CPU has no L3 cache domain, or which has no ROCm library or
ethtool support, used to fail a readiness check over hardware it was never
going to have. Monitoring that alerts on the status exit code will stop firing
on such hosts; a genuine initialization failure still exits non-zero, and
disabled,unsupportedandfailedare now three answers rather than two.
(#1102, #1103) - GPU: a Tegra SoC's integrated GPU no longer records utilization or PCIe link
metrics. NVML answers both queries there and answers wrongly — a hard-coded
0% utilization indistinguishable from a genuinely idle GPU once it is a row in
an archive, and a placeholder generation and width for a link an iGPU does not
have.gpu_sm_utilization,gpu_dram_bandwidth_utilizationand the PCIe
series are absent on those parts now rather than present and wrong. (#1108) - Viewer/MCP: a query against a
.rezarchive is 12.8x faster. A query opens
only the tables it touches, where before it materialized the whole archive and
spent 91% of its wall time opening tables it never read..rezis now at or
ahead of a single parquet holding the same window. No format change — existing
archives get this on read. (#1107)
Added
- Recorder: several endpoints record into one
.rez.rezolus record --endpoint http://web-01:4241 --endpoint http://web-02:4241 -o fleet.rezwrites one
archive holding a recording per host, where before more than one endpoint
silently demoted the output to parquet. Both arms of an A/B can now be
captured concurrently rather than in sequence, so they share their background
load instead of differing in it as well as in the experiment. (#1109) - Packaging:
.rpmpackages for Enterprise Linux 10, and the install script now
accepts Rocky Linux, AlmaLinux, RHEL, CentOS Stream and Oracle Linux at major
9 or 10.curl -fsSL https://install.rezolus.com | sudo bashfailed outright
on EL10 before. (#1113)
Fixed
- Recording: reading a
.rezwhile it is being written no longer panics when
retention evicts a table mid-query. A quiet sampler's only rows could be
dropped between the probe that named its table and the query that opened it,
crashing the viewer or MCP against a hindsight rolling buffer — which is the
one thing that buffer exists to be read as. A table that has gone is reported
absent, the same as one that was never there. (#1112) - Recording: a Prometheus source that emits timestamps no longer writes
acquisition windows anchored decades before the rows holding them. The
optional trailing timestamp in the exposition format is a federation staleness
marker, not a claim about when we read; taking it at face value gave a row
recorded in 2026 a window one second after the Unix epoch — a ~56-year error
in the operandrate()prices its uncertainty from, nonsense rather than
conservative. The window is our fetch. Any pushgateway or federation exporter
produced this, on the parquet path as well as.rez. (#1111)