High cardinality, no problem: Quiver held 100M series on 2 cores and 2 GB

10k samples/sfixed; only the number of series changes
2 cores / 2 GBper engine, the same for all three
AWS c8a.2xlargeAMD EPYC, x86, one host
Kubernetes metrics68 series per pod, Prometheus labels

On the same limits, VictoriaMetrics ran out of memory at 5M series and InfluxDB at 1M. Quiver's engine stayed under 100 MiB the whole way.

QuiverVictoriaMetricsInfluxDB
2 cores / 2 GB per engine · 10,000 samples a second
QuiverVictoriaMetricsInfluxDB
Largest size held100Mall we tested, no limit reached3Mout of memory at 5M500kout of memory at 1M
Engine memory at 1M series81 MiB492 MiB–out of memory
Engine memory at its largest size95 MiBat 100M1.23 GBat 3M, peak 1.92 GB949 MiBat 500k
New series taken in, at its largest size1.60M/ssamples a second146k/s100k/s
Why telemetry is different

Telemetry is written all day in huge volumes, never changed once written, and read rarely: AI agents query it in short bursts during an investigation, not dashboards refreshing all day. VictoriaMetrics and InfluxDB register every series in an index so a dashboard can find it in a millisecond; that index, and the memory it needs, grows with every new label combination. Quiver keeps no series index, so the number of series barely changes its cost.

Series held

Every size we ran

Each tile is one run. Kept up: took in every series, held the write rate, and every sample checked came back exactly. Quiver passed every size up to 100M, the largest we tested, without reaching its limit. For VictoriaMetrics and InfluxDB we ran the last size that passed and the first that failed.

Quiver
100k seriesKept up79 MiB · 0.04 cores
1M seriesKept up81 MiB · 0.04 cores
10M seriesKept up81 MiB · 0.06 cores
100M seriesKept up95 MiB · 0.09 cores
VictoriaMetrics
100k seriesKept up106 MiB · 0.02 cores
1M seriesKept up492 MiB · 0.02 cores
3M seriesKept up1.23 GB · 0.04 cores
5M seriesOut of memoryafter 45 s · about 4.0M created
InfluxDB
100k seriesKept up309 MiB · 0.13 cores
500k seriesKept up949 MiB · 0.16 cores
1M seriesOut of memoryafter 50 s · 1M created
Memory

Quiver's memory doesn't grow with the number of series

Engine memory while ingesting, by unique series

QuiverVictoriaMetricsInfluxDB
0 GB0.5 GB1 GB1.5 GB2 GB100k1M10M100Munique series (log scale)memory cap per engine: 2 GBQuiver, 100k series: 79 MiBQuiver, 1M series: 81 MiBQuiver, 10M series: 81 MiBQuiver, 100M series: 95 MiBVictoriaMetrics, 100k series: 106 MiBVictoriaMetrics, 1M series: 492 MiBVictoriaMetrics, 3M series: 1.23 GBVictoriaMetrics, 5M series: out of memoryInfluxDB, 100k series: 309 MiBInfluxDB, 500k series: 949 MiBInfluxDB, 1M series: out of memoryQuiver: 95 MiB at 100MVictoriaMetrics: out of memory at 5MInfluxDB: out of memory at 1M

Engine memory while ingesting. A cross marks an engine killed at its 2 GB cap. Not included: Garage, the S3-compatible object store Quiver writes to (20–325 MiB), and Telegraf, InfluxDB's remote-write adapter. Both are in Addendum A.

Never-repeating series

Pods replaced every 2 minutes: Quiver stays flat

200,000 series active at once, every pod renamed every 2 minutes, as deploys and autoscaling do: 1.4M distinct series in 12 minutes.

Memory per quarter, incl. Garage / Telegraf

QuiverVictoriaMetricsInfluxDB
0.0 GB0.5 GB1.0 GB1.5 GB2.0 GBQ1Q2Q3Q4quarter of the 12-minute live phaseQuiver, quarter 1: 0.1 GBQuiver, quarter 2: 0.1 GBQuiver, quarter 3: 0.1 GBQuiver, quarter 4: 0.1 GBVictoriaMetrics, quarter 1: 0.8 GBVictoriaMetrics, quarter 2: 0.6 GBVictoriaMetrics, quarter 3: 0.8 GBVictoriaMetrics, quarter 4: 1.0 GBInfluxDB, quarter 1: 0.9 GBInfluxDB, quarter 2: 1.0 GBInfluxDB, quarter 3: 1.2 GBInfluxDB, quarter 4: 1.4 GB
Memory, first to last quarter
Quiver+3%99 MiB average
VictoriaMetrics+25%835 MiB average
InfluxDB+54%1.20 GB average
Agent queries

Built for agents, not dashboards

VictoriaMetrics answers from its index in under a millisecond. Quiver scans the time range instead: 7 ms to 1.2 s.

That matters to a dashboard refreshing dozens of panels. It doesn't to an agent: each query sits between LLM calls that take seconds, so a few hundred milliseconds is lost in the wait. Skipping the index is also why Quiver doesn't mind cardinality.

QueryQuiverVictoria­MetricsInfluxDB
Memory of one pod (last 15 min)2810.22.5
CPU rate of one container (last 1 h)2150.12.7
Request rate per pod of one service (last 1 h)1,0470.4231
Top 5 pods by CPU in one namespace (now)1020.415.2
5xx error ratio of one service (last 1 h)1,1860.3326
Average memory of one pod (last 1 h)2940.12.3
Which pods one service has (last 1 h)6.70.12.2

Typical (p50) response in milliseconds at 100k series, the size all three hold. 20 runs each; every answer checked against the true value. Similar with ingest running.

How we tested

Method

  • Data: Kubernetes-style metrics, 68 series per pod with cluster, namespace, service, pod, node and job labels. Identical for every engine.
  • A fixed 10,000 samples a second at every size; only the number of series changes. Each run writes the history (every series at least twice), then a live phase, then checks.
  • Ingest: Quiver over Arrow Flight, VictoriaMetrics over Prometheus remote write, InfluxDB through Telegraf (InfluxDB 2 has no remote-write endpoint).
  • Limits: 2 cores / 2 GB per engine, no swap, so going over the cap means being killed. Garage 1 core / 1 GB; Telegraf 2 cores / 2 GB.
  • Not compared: storage size (Quiver's compaction did not run in these short tests), histogram percentile queries, clusters and paid editions.
Addenda

Every run and versions

AEvery run

EngineUnique seriesEach series written everyResultCPU, cores (incl. Garage / Telegraf)Engine memoryMemory incl. Garage / TelegrafEngine memory peakSeries creation, samples/sSample count
Quiver100k10 sKept up0.0479 MiB125 MiB581 MiB2.21Mexact
Quiver1M100 sKept up0.0481 MiB101 MiB539 MiB2.65Mexact
Quiver10M1,000 sKept up0.0681 MiB116 MiB204 MiB1.29Mexact
Quiver100M10,000 sKept up0.0995 MiB420 MiB1.15 GB1.60Mexact
VictoriaMetrics100k10 sKept up0.02106 MiB106 MiB348 MiB2.35Mexact
VictoriaMetrics1M100 sKept up0.02492 MiB492 MiB1.06 GB487kexact
VictoriaMetrics3M300 sKept up0.041.23 GB1.23 GB1.92 GB146kexact
VictoriaMetrics5M500 sOut of memory–––2.00 GB––
InfluxDB100k10 sKept up0.13309 MiB441 MiB526 MiB164kexact
InfluxDB500k50 sKept up0.16949 MiB1.06 GB1.92 GB100kexact
InfluxDB1M–Out of memory–––1.91 GB––

Pods replaced every 2 minutes, per quarter of the live phase:

EngineQ1 CPUQ2 CPUQ3 CPUQ4 CPUQ1 memoryQ2 memoryQ3 memoryQ4 memoryEngine memory peakSample count
Quiver0.030.040.050.05103 MiB96 MiB95 MiB106 MiB180 MiBexact
VictoriaMetrics0.060.060.060.06783 MiB664 MiB863 MiB979 MiB1.06 GBexact
InfluxDB0.220.230.200.22944 MiB0.99 GB1.18 GB1.42 GB1.82 GBexact

BVersions

Machine
AWS us-east-1, c8a.2xlarge (8 vCPU AMD EPYC 9R45, 16 GiB), 200 GB gp3 (3,000 IOPS, 250 MB/s), Ubuntu 24.04, Docker 29.9.3
Quiver
core0 build 35b4016, single node
Object storage
Garage v2.5.0, single node, same disk
VictoriaMetrics
v1.155.0 single-node, open source (Apache 2.0), default settings
InfluxDB
2.9.1 OSS (MIT), default settings; Telegraf 1.40.3 (MIT), prometheusremotewrite parser, metric_version 1