feat: add s3 variant

README.md

@@ -159,7 +159,7 @@ emits benchstat output) can be pointed at any `addr` from `systems.yml` — see
   in `queries/native/`; they measure the same retrieval work, not identical
   semantics. Queries with no equivalent report `n/a`.
 - **Datasets.** Metrics are live `node_exporter` series multiplied to 10 hosts by
-  vmagent during a prewarm (`PREWARM` env, default 90s); some engines have PromQL
+  vmagent during a prewarm (`PREWARM` env, default 30s); some engines have PromQL
   dialect gaps (e.g. GreptimeDB rejects `topk` and bare `{__name__=~...}`),
   reported as `ERR`/empty. Logs match
   the loghub suite when `LOGHUB_DIR` is set, else synthetic. Traces are

cmd/benchctl/main.go

@@ -54,6 +54,8 @@ func run(args []string) error {
 		return e.Ingest(must(rest, 0, "ingest <metrics|logs|traces>"))
 	case "query":
 		return e.Query(must(rest, 0, "query <metrics|logs|traces> [runs]"), runs(rest, 1, e.Runs), 3)
+	case "check":
+		return e.Check(must(rest, 0, "check <metrics|logs|traces>"))
 	case "collect":
 		return e.Collect()
 	case "report":
@@ -92,6 +94,9 @@ func benchLane(e *bench.Env, lane string, n int) error {
 	if err := e.Query(lane, n, 3); err != nil {
 		return err
 	}
+	if err := e.Check(lane); err != nil {
+		return err
+	}
 	return e.Collect()
 }
 
@@ -101,6 +106,7 @@ func usage() {
   up      [metrics|logs|traces|all]    build + start a lane (default all)
   ingest  <metrics|logs|traces>        push canonical dataset (fan-out)
   query   <metrics|logs|traces> [runs] replay suite against every engine
+  check   <metrics|logs|traces>        verify query results (count vs reference)
   collect                              capture storage + memory footprint
   report                               render results/REPORT.md
   bench   <metrics|logs|traces> [runs] up→ingest→settle→query→collect

compose/base.yml

@@ -51,7 +51,7 @@ services:
     entrypoint: ["/bin/sh", "-c"]
     command:
       - |
-        for b in mimir loki tempo greptimedb; do
+        for b in mimir loki tempo greptimedb oteldb-s3; do
           aws --endpoint-url http://fs:9000 s3 mb "s3://$$b" 2>/dev/null || true
         done
         echo "buckets ready"

compose/metrics.yml

@@ -28,6 +28,7 @@ services:
       - -promscrape.config=/etc/vmagent/scrape.yml
       - -remoteWrite.url=http://oteldb:19291/api/v1/write
       - -remoteWrite.url=http://oteldb-ch:19291/api/v1/write
+      - -remoteWrite.url=http://oteldb-s3:19291/api/v1/write
       - -remoteWrite.url=http://prometheus:9090/api/v1/write
       - -remoteWrite.url=http://victoriametrics:8428/api/v1/write
       - -remoteWrite.url=http://mimir:9009/api/v1/push
@@ -77,9 +78,11 @@ services:
     ports: ["127.0.0.1:9009:9009"]
 
   # GreptimeDB standalone, object storage on go-faster/fs S3 (bucket: greptimedb).
+  # In [logs]/[traces] too: it ingests the OTLP log/trace streams into otel_logs /
+  # opentelemetry_traces and is queried over SQL (systems.yml, queries/native/greptimedb.yml).
   greptimedb:
-    image: greptime/greptimedb:v0.11.0
-    profiles: [metrics]
+    image: greptime/greptimedb:v1.1.1
+    profiles: [metrics, logs, traces]
     networks: [bench]
     depends_on:
       fs-init: { condition: service_completed_successfully }

compose/shared.yml

@@ -7,6 +7,7 @@
 
 volumes:
   oteldb-data:
+  oteldb-s3-data:
   clickhouse-data:
   clickhouse-oteldb-data:
 
@@ -38,6 +39,32 @@ services:
       - "127.0.0.1:9010:9010"   # pprof
       - "127.0.0.1:19291:19291" # prometheus remote-write ingest
 
+  # oteldb, S3 variant: same image and --embedded, but the embedded storage
+  # engine's object-store backend is the shared go-faster/fs S3 (bucket oteldb-s3),
+  # putting oteldb in the apples-to-apples object-store group. Only the WAL is local.
+  oteldb-s3:
+    image: ${OTELDB_IMAGE:-ghcr.io/oteldb/oteldb:v0.41.0}
+    command: ["--config=/etc/otel/cfg.yml", "--embedded"]
+    networks: [bench]
+    depends_on:
+      fs-init: { condition: service_completed_successfully }
+    volumes:
+      - ../config/oteldb/oteldb-s3.yml:/etc/otel/cfg.yml:ro
+      - oteldb-s3-data:/data
+    environment:
+      - OTEL_LOG_LEVEL=info
+      - OTEL_METRICS_EXPORTER=none
+      - OTEL_LOGS_EXPORTER=none
+      - OTEL_TRACES_EXPORTER=none
+      - PPROF_ADDR=0.0.0.0:9010
+      - GOMAXPROCS=${BENCH_CPUS:-}
+      - GOMEMLIMIT=${BENCH_MEMLIMIT:-}
+    ports:
+      - "127.0.0.1:9093:9090"   # promql
+      - "127.0.0.1:3103:3100"   # logql
+      - "127.0.0.1:3203:3200"   # traceql
+      - "127.0.0.1:9012:9010"   # pprof
+
   # oteldb, ClickHouse-backed variant: same image WITHOUT --embedded, serving all
   # three signals from oteldb's chstorage on a dedicated ClickHouse (so its disk
   # footprint is isolated from gigapipe's). Distinct host ports from the embedded

config/otelcol/logs.yml

@@ -17,9 +17,19 @@ exporters:
   otlp/oteldb-ch:
     endpoint: oteldb-ch:4317
     tls: { insecure: true }
+  otlp/oteldb-s3:
+    endpoint: oteldb-s3:4317
+    tls: { insecure: true }
   otlphttp/victorialogs:
     logs_endpoint: http://victorialogs:9428/insert/opentelemetry/v1/logs
     tls: { insecure: true }
+  otlphttp/greptimedb:
+    # GreptimeDB OTLP logs: default (no pipeline) -> structured table with a
+    # fulltext-indexed `body`, `severity_text`/`severity_number`, and JSON
+    # log/resource attributes. Table selected by the x-greptime-log-table-name header.
+    logs_endpoint: http://greptimedb:4000/v1/otlp/v1/logs
+    headers: { x-greptime-log-table-name: otel_logs }
+    tls: { insecure: true }
   loki/loki:
     endpoint: http://loki:3100/loki/api/v1/push
     tls: { insecure: true }
@@ -35,6 +45,8 @@ service:
       exporters:
         - otlp/oteldb
         - otlp/oteldb-ch
+        - otlp/oteldb-s3
         - otlphttp/victorialogs
+        - otlphttp/greptimedb
         - loki/loki
         - loki/gigapipe

config/otelcol/traces.yml

@@ -15,12 +15,24 @@ exporters:
   otlp/oteldb-ch:
     endpoint: oteldb-ch:4317
     tls: { insecure: true }
+  otlp/oteldb-s3:
+    endpoint: oteldb-s3:4317
+    tls: { insecure: true }
   otlp/tempo:
     endpoint: tempo:4317
     tls: { insecure: true }
   otlphttp/victoriatraces:
     traces_endpoint: http://victoriatraces:10428/insert/opentelemetry/v1/traces
     tls: { insecure: true }
+  otlphttp/greptimedb:
+    # GreptimeDB OTLP traces -> table opentelemetry_traces (columns span_name,
+    # duration_nano, parent_span_id, span_attributes/resource_attributes JSON),
+    # queried via SQL. v1.1.x requires the built-in trace pipeline header.
+    traces_endpoint: http://greptimedb:4000/v1/otlp/v1/traces
+    headers:
+      x-greptime-trace-table-name: opentelemetry_traces
+      x-greptime-pipeline-name: greptime_trace_v1
+    tls: { insecure: true }
   otlp/gigapipe:
     endpoint: gigapipe:4317
     tls: { insecure: true }
@@ -33,6 +45,8 @@ service:
       exporters:
         - otlp/oteldb
         - otlp/oteldb-ch
+        - otlp/oteldb-s3
         - otlp/tempo
         - otlphttp/victoriatraces
+        - otlphttp/greptimedb
         - otlp/gigapipe

config/oteldb/oteldb-s3.yml

@@ -0,0 +1,58 @@
+# oteldb under test, S3 variant. Same embedded go-faster/storage engine as
+# config/oteldb/oteldb.yml, but its object-store backend is the shared
+# go-faster/fs S3 server (bucket: oteldb-s3) instead of a local `file` volume —
+# the apples-to-apples object-store substrate Mimir/Loki/Tempo/GreptimeDB use.
+# Only the WAL is local (S3 is stateless); see storage.s3 below.
+storage:
+  backend: s3
+  # Local write-ahead log (the only on-disk state for the s3 backend).
+  wal_dir: /data/wal
+  s3:
+    bucket: oteldb-s3
+    endpoint: http://fs:9000
+    region: us-east-1
+    force_path_style: true       # go-faster/fs serves path-style (endpoint/bucket/key)
+    access_key_id: bench
+    secret_access_key: benchbench
+  # Parallelize LogQL result materialization across workers, like the file variant.
+  log_query_parallelism: 8
+  # Flush head data to S3 on this interval so a short bench actually exercises (and
+  # measures the footprint of) the object store rather than serving only from the
+  # in-memory head. Default (0) is size-triggered only — a one-shot ingest would
+  # never reach S3.
+  flush_interval: 10s
+
+prometheus:                     # PromQL HTTP API (Prometheus-compatible)
+  bind: 0.0.0.0:9090
+  max_samples: 1_000_000
+  timeout: 1m
+  enable_negative_offset: true
+loki:                           # LogQL HTTP API (Loki-compatible)
+  bind: 0.0.0.0:3100
+tempo:                          # TraceQL HTTP API (Tempo-compatible)
+  bind: 0.0.0.0:3200
+health_check:
+  bind: 0.0.0.0:13133
+
+# Ingest receivers: Prometheus remote-write for metrics, OTLP for logs + traces.
+# Under --embedded the oteldbexporter writes to the embedded storage engine, so
+# its DSN is ignored — but every pipeline must still name an exporter.
+otelcol:
+  receivers:
+    prometheusremotewrite:
+      endpoint: 0.0.0.0:19291
+      time_threshold: 100000h
+    otlp:
+      protocols:
+        grpc:
+          endpoint: 0.0.0.0:4317
+        http:
+          endpoint: 0.0.0.0:4318
+  exporters:
+    oteldbexporter:
+      dsn: clickhouse://default:default@clickhouse:9000   # ignored under --embedded
+  service:
+    pipelines:
+      metrics: { receivers: [prometheusremotewrite], exporters: [oteldbexporter] }
+      logs:    { receivers: [otlp], exporters: [oteldbexporter] }
+      traces:  { receivers: [otlp], exporters: [oteldbexporter] }