System fixes from #904 (excluding /playground)

byconity/benchmark.sh

@@ -7,4 +7,10 @@ export BENCH_DURABLE=yes
 # dependency, so the worst-case cold start is several minutes; the
 # lib's 300s default has timed out before server is up.
 export BENCH_CHECK_TIMEOUT=1200
+# After firecracker snapshot+restore the cluster's
+# internal connections (brpc/gossip) are stale; ./start's
+# shallow health probe doesn't notice and short-circuits.
+# Tell the playground agent to ./stop the cluster before
+# ./start so the next bring-up is from a clean state.
+export PLAYGROUND_RESTART_AFTER_RESTORE_SNAPSHOT=yes
 exec ../lib/benchmark-common.sh

cedardb-parquet/start

@@ -28,10 +28,15 @@ if ! sudo docker run -d --rm -p 5432:5432 \
     exit 1
 fi
 
-for _ in $(seq 1 60); do
+# First-boot initdb inside the container takes well over a minute
+# (observed ~90-120 s of "Fixing permissions"/"Setting up database
+# directory" before postgres actually listens). Give it 10 min —
+# pg_isready exits fast once the daemon is up, so this only
+# matters in the failure path.
+for _ in $(seq 1 600); do
     pg_isready -h localhost --dbname postgres -U postgres >/dev/null 2>&1 && exit 0
     sleep 1
 done
-echo "cedardb did not become ready in 60 s; container logs:" >&2
+echo "cedardb did not become ready in 600 s; container logs:" >&2
 sudo docker logs cedardb 2>&1 | tail -40 >&2 || true
 exit 1

cedardb/start

@@ -30,10 +30,15 @@ if ! sudo docker run -d --rm -p 5432:5432 \
     exit 1
 fi
 
-for _ in $(seq 1 60); do
+# First-boot initdb inside the container can run for well over a
+# minute (observed ~90-120 s of "Fixing permissions"/"Setting up
+# database directory" before postgres actually listens). Older
+# 60 s budget bailed during that phase. Give it 5 min — pg_isready
+# exits fast once the daemon is up so this only matters on failure.
+for _ in $(seq 1 600); do
     pg_isready -h localhost --dbname postgres -U postgres >/dev/null 2>&1 && exit 0
     sleep 1
 done
-echo "cedardb did not become ready in 60 s; container logs:" >&2
+echo "cedardb did not become ready in 600 s; container logs:" >&2
 sudo docker logs cedardb 2>&1 | tail -40 >&2 || true
 exit 1

chdb-dataframe/benchmark.sh

@@ -2,4 +2,9 @@
 # Thin shim — actual flow is in lib/benchmark-common.sh.
 export BENCH_DOWNLOAD_SCRIPT="download-hits-parquet-single"
 export BENCH_DURABLE=no
+# Skip the pre-snapshot ./stop+./start cycle: the loaded
+# state lives only in the daemon's process memory (in-process
+# DataFrame, JVM heap caches) and stopping wipes it. The
+# playground agent reads this and snapshots the running daemon.
+export PLAYGROUND_SKIP_RESTART_BEFORE_SNAPSHOT=yes
 exec ../lib/benchmark-common.sh

clickhouse-web/create.sql

@@ -108,5 +108,5 @@ ATTACH TABLE hits UUID 'c449dfbf-ba06-4d13-abec-8396559eb955'
     PRIMARY KEY (CounterID, EventDate, UserID, EventTime, WatchID)
 )
 ENGINE = MergeTree
-SETTINGS disk = disk(type = cache, path = '/dev/shm/clickhouse/', max_size_ratio_to_total_space = 0.9,
+SETTINGS disk = disk(type = cache, path = '/var/lib/clickhouse/caches/web/', max_size_ratio_to_total_space = 0.9,
          disk = disk(type = web, endpoint = 'https://clickhouse-public-datasets.s3.amazonaws.com/web/'));

clickhouse-web/install

@@ -10,6 +10,18 @@ if [ ! -x /usr/bin/clickhouse ]; then
     sudo ./clickhouse install --noninteractive
 fi
 
-# Cache directory used by the web disk.
-sudo mkdir -p /dev/shm/clickhouse
-sudo chown clickhouse:clickhouse /dev/shm/clickhouse
+# Cache directory used by the web disk. ClickHouse rejects any
+# filesystem-cache path outside /var/lib/clickhouse/caches/ with
+# BAD_ARGUMENTS at CREATE TABLE time, but we still want the actual
+# bytes to live in tmpfs (/dev/shm) for the speed: cold queries
+# pull ~1 GB on first run and tmpfs avoids touching the host SSD.
+#
+# Newer ClickHouse versions canonicalise the path before the policy
+# check, so the older symlink trick (caches/web → /dev/shm/...) is
+# rejected with BAD_ARGUMENTS. Bind-mount tmpfs at the
+# policy-acceptable path instead — to CH the cache dir *is*
+# /var/lib/clickhouse/caches/web with no symlink to resolve.
+sudo mkdir -p /dev/shm/clickhouse /var/lib/clickhouse/caches/web
+sudo chown clickhouse:clickhouse /dev/shm/clickhouse /var/lib/clickhouse/caches/web
+sudo mount --bind /dev/shm/clickhouse /var/lib/clickhouse/caches/web
+sudo chown clickhouse:clickhouse /var/lib/clickhouse/caches/web

daft-parquet-partitioned/benchmark.sh

@@ -2,4 +2,9 @@
 # Thin shim — actual flow is in lib/benchmark-common.sh.
 export BENCH_DOWNLOAD_SCRIPT="download-hits-parquet-partitioned"
 export BENCH_DURABLE=no
+# Skip the pre-snapshot ./stop+./start cycle: the loaded
+# state lives only in the daemon's process memory (in-process
+# DataFrame, JVM heap caches) and stopping wipes it. The
+# playground agent reads this and snapshots the running daemon.
+export PLAYGROUND_SKIP_RESTART_BEFORE_SNAPSHOT=yes
 exec ../lib/benchmark-common.sh

daft-parquet/benchmark.sh

@@ -2,4 +2,9 @@
 # Thin shim — actual flow is in lib/benchmark-common.sh.
 export BENCH_DOWNLOAD_SCRIPT="download-hits-parquet-single"
 export BENCH_DURABLE=no
+# Skip the pre-snapshot ./stop+./start cycle: the loaded
+# state lives only in the daemon's process memory (in-process
+# DataFrame, JVM heap caches) and stopping wipes it. The
+# playground agent reads this and snapshots the running daemon.
+export PLAYGROUND_SKIP_RESTART_BEFORE_SNAPSHOT=yes
 exec ../lib/benchmark-common.sh

druid/load

@@ -11,8 +11,15 @@ DRUID_DIR="apache-druid-${VERSION}"
 # datasource).
 "./${DRUID_DIR}/bin/post-index-task" --file ingest.json --url http://localhost:8081 || true
 
-# Wait until the hits datasource is queryable.
-for _ in $(seq 1 600); do
+# Wait until the hits datasource is queryable. Druid's index task can
+# legitimately take hours on a 16 GiB VM; budget 4 h here, and fail
+# loudly if hits still isn't queryable so the agent doesn't take a
+# snapshot of a half-ingested datasource (which would otherwise look
+# "snapshotted" but every query returns
+#   druidException ... Object 'hits' not found
+# at runtime).
+cnt=""
+for _ in $(seq 1 2880); do  # 2880 * 5s = 4 h
     cnt=$(curl -sf -XPOST -H'Content-Type: application/json' \
         http://localhost:8888/druid/v2/sql/ \
         -d '{"query": "SELECT COUNT(*) FROM hits"}' 2>/dev/null \
@@ -22,6 +29,13 @@ for _ in $(seq 1 600); do
     fi
     sleep 5
 done
+if [ -z "$cnt" ] || [ "$cnt" -le 0 ]; then
+    echo "druid: hits datasource still not queryable after 4 h; ingestion" >&2
+    echo "did not finish. Dumping recent task list for diagnosis:" >&2
+    curl -sS http://localhost:8081/druid/indexer/v1/tasks 2>&1 | head -c 2000 >&2
+    exit 1
+fi
+echo "druid: hits has $cnt rows after ingestion"
 
 rm -f hits.tsv
 sync

duckdb-dataframe/benchmark.sh

@@ -2,4 +2,9 @@
 # Thin shim — actual flow is in lib/benchmark-common.sh.
 export BENCH_DOWNLOAD_SCRIPT="download-hits-parquet-single"
 export BENCH_DURABLE=no
+# Skip the pre-snapshot ./stop+./start cycle: the loaded
+# state lives only in the daemon's process memory (in-process
+# DataFrame, JVM heap caches) and stopping wipes it. The
+# playground agent reads this and snapshots the running daemon.
+export PLAYGROUND_SKIP_RESTART_BEFORE_SNAPSHOT=yes
 exec ../lib/benchmark-common.sh

firebolt-parquet-partitioned/benchmark.sh

@@ -1,49 +1,6 @@
 #!/bin/bash
-
-# Download the partitioned hits parquet files
-echo "Downloading dataset..."
-rm -rf data
-../lib/download-hits-parquet-partitioned data
-
-# Start the container
-sudo apt-get install -y docker.io jq
-sudo docker run -dit --name firebolt-core --rm \
-    --ulimit memlock=8589934592:8589934592 \
-    --security-opt seccomp=unconfined \
-    -p 127.0.0.1:3473:3473 \
-    -v /firebolt-core/volume \
-    -v ./data/:/firebolt-core/clickbench \
-    ghcr.io/firebolt-db/firebolt-core:preview-rc
-
-# See firebolt/benchmark.sh — the old curl-and-break pattern accepted the
-# "Cluster not yet healthy" JSON error body as success.
-for _ in {1..600}
-do
-    if curl -sS "http://localhost:3473/" \
-            --data-binary "SELECT 'Firebolt is ready';" 2>/dev/null \
-            | grep -q "Firebolt is ready"; then
-        break
-    fi
-    sleep 1
-done
-
-# Create the database and external table
-echo "Creating external table..."
-curl -sS "http://localhost:3473/?enable_multi_query_requests=true" --data-binary "DROP DATABASE IF EXISTS clickbench;CREATE DATABASE clickbench;"
-curl -sS "http://localhost:3473/?database=clickbench&enable_multi_query_requests=true" --data-binary @create.sql
-
-# Print statistics
-DATA_SIZE=$(du -bcs data/hits_*.parquet 2>/dev/null | grep total | awk '{print $1}')
-if [ -z "$DATA_SIZE" ]; then
-    DATA_SIZE=$(du -cs data/hits_*.parquet | grep total | awk '{print $1}')
-fi
-echo "Load time: 0"
-echo "Data size: $DATA_SIZE"
-
-# Run the benchmark
-echo "Running the benchmark..."
-./run.sh
-
-# Stop the container and remove the data
-sudo docker container stop firebolt-core
-rm -rf data
+# Thin shim — actual flow is in lib/benchmark-common.sh.
+export BENCH_DOWNLOAD_SCRIPT="download-hits-parquet-partitioned"
+export BENCH_DURABLE=no
+export BENCH_RESTARTABLE=no
+exec ../lib/benchmark-common.sh

firebolt-parquet-partitioned/check

@@ -0,0 +1,7 @@
+#!/bin/bash
+set -e
+
+# Firebolt-core's HTTP port answers immediately but may return a
+# cluster-not-ready JSON error at HTTP 200. Test for an actual result.
+curl -sSf --max-time 5 'http://localhost:3473/' \
+    --data-binary 'SELECT 1;' 2>/dev/null | grep -q '^1'

firebolt-parquet-partitioned/data-size

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -e
+
+# Firebolt-core writes its database state under /firebolt-core/volume
+# inside the container, which we bind-mount to ./fb-volume on the host.
+du -bcs fb-volume 2>/dev/null | awk '/total$/ { print $1 }'

firebolt-parquet-partitioned/install

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -eu
+
+sudo apt-get update -y
+sudo apt-get install -y docker.io jq
+sudo docker pull ghcr.io/firebolt-db/firebolt-core:preview-rc

firebolt-parquet-partitioned/load

@@ -0,0 +1,20 @@
+#!/bin/bash
+set -eu
+
+# Partitioned-parquet variant: stage hits_*.parquet under ./data so
+# the container sees them at /firebolt-core/clickbench/*.parquet;
+# create.sql declares an external table with FROM PATTERN that
+# matches the glob.
+mkdir -p data
+shopt -s nullglob
+for f in hits_*.parquet; do
+    mv -f "$f" "data/$f"
+done
+shopt -u nullglob
+
+curl -sSf 'http://localhost:3473/?enable_multi_query_requests=true' \
+    --data-binary 'DROP DATABASE IF EXISTS clickbench;CREATE DATABASE clickbench;'
+curl -sSf 'http://localhost:3473/?database=clickbench&enable_multi_query_requests=true' \
+    --data-binary @create.sql
+
+sync

firebolt-parquet-partitioned/query

@@ -0,0 +1,28 @@
+#!/bin/bash
+# Reads a SQL query from stdin, runs it against the firebolt-core
+# container via /?database=clickbench.
+# Stdout: query result (firebolt's JSON_Compact format).
+# Stderr: query runtime in fractional seconds on the last line,
+#         pulled from the response's `.statistics.elapsed`.
+# Exit non-zero on error.
+set -e
+
+query=$(cat)
+
+# Result + sub-result caches off so timings are real; output_format
+# matches what firebolt's run.sh uses for the public benchmark.
+PARAMS='database=clickbench&enable_result_cache=false&enable_subresult_cache=false&enable_scan_cache=false&output_format=JSON_Compact'
+
+resp=$(curl -sS --max-time 600 "http://localhost:3473/?${PARAMS}" \
+            --data-binary "$query")
+
+# Firebolt returns a JSON object whether the query succeeded or not.
+# A failed query has an "errors" key; a successful one carries
+# "data" + "statistics".
+if printf '%s' "$resp" | jq -e '.errors' >/dev/null 2>&1; then
+    printf '%s\n' "$resp" >&2
+    exit 1
+fi
+
+printf '%s\n' "$resp"
+printf '%s\n' "$resp" | jq -r '.statistics.elapsed' >&2

firebolt-parquet-partitioned/start

@@ -0,0 +1,76 @@
+#!/bin/bash
+set -eu
+
+# Idempotent: if firebolt-core already answers SELECT 1, do nothing.
+if curl -sS --max-time 5 'http://localhost:3473/' \
+        --data-binary 'SELECT 1;' 2>/dev/null | grep -q '^1'; then
+    exit 0
+fi
+
+mkdir -p data fb-volume
+# firebolt-core runs as UID/GID 1111 inside the container and refuses
+# to start if its data dir is not writeable by that uid (the engine
+# self-checks and aborts with "directory ... is not readable or
+# writeable by the Firebolt Core process"). Set the host-side
+# ownership accordingly so the bind-mounted dir is usable.
+sudo chown 1111:1111 fb-volume
+
+# If the container exists (stopped from a prior agent pre-snapshot
+# cycle), just start it back — the data lives on the bind-mounted
+# fb-volume below, so the previously-created `clickbench` database
+# is still there. Otherwise create the container fresh.
+if sudo docker ps -a --format '{{.Names}}' | grep -qx firebolt-core; then
+    sudo docker start firebolt-core >/dev/null
+else
+    # `firebolt-core` is the public self-hosted image. Container needs
+    # memlock 8 GiB and seccomp unconfined per upstream's run docs.
+    # /firebolt-core/clickbench: parquet source (read at load time).
+    # /firebolt-core/volume:     engine data directory (must persist
+    #                            across the agent's pre-snapshot
+    #                            stop+start cycle or the snapshot
+    #                            ships an empty DB).
+    # The agent stages partitioned parquet at $PWD as symlinks pointing
+    # at /opt/clickbench/datasets_ro/hits_partitioned/hits_N.parquet
+    # (an absolute host-VM path). `./load` then `mv`s those symlinks
+    # into data/, but inside the container the absolute target is
+    # unreachable and the symlinks dangle — load reports success but
+    # the external table sees zero files and every query returns
+    # rows_read=0. Bind-mount the dataset disk into the container at
+    # the same absolute path so the symlinks resolve.
+    sudo docker run -dit --name firebolt-core \
+        --ulimit memlock=8589934592:8589934592 \
+        --security-opt seccomp=unconfined \
+        -p 127.0.0.1:3473:3473 \
+        -v "$(pwd)/fb-volume:/firebolt-core/volume" \
+        -v "$(pwd)/data:/firebolt-core/clickbench" \
+        -v "/opt/clickbench/datasets_ro:/opt/clickbench/datasets_ro:ro" \
+        ghcr.io/firebolt-db/firebolt-core:preview-rc >/dev/null
+fi
+
+# Wait for the cluster to be "actually" ready. firebolt-core's HTTP
+# port comes up immediately but returns
+#   {"errors":[{"description":"Cluster not yet healthy: ..."}]}
+# at HTTP 200 until the engine threads have warmed; bench against a
+# sentinel string instead of HTTP status to avoid that trap.
+for _ in $(seq 1 600); do
+    if curl -sS --max-time 5 'http://localhost:3473/' \
+            --data-binary "SELECT 'firebolt-ready';" 2>/dev/null \
+            | grep -q 'firebolt-ready'; then
+        exit 0
+    fi
+    sleep 1
+done
+{
+    echo "firebolt-core did not become healthy in 10 min"
+    echo "=== docker ps -a ==="
+    sudo docker ps -a 2>&1
+    echo "=== docker inspect firebolt-core (state) ==="
+    sudo docker inspect firebolt-core --format '{{json .State}}' 2>&1
+    echo "=== docker logs firebolt-core --tail 50 ==="
+    sudo docker logs firebolt-core --tail 50 2>&1
+    echo "=== curl http://localhost:3473/ ==="
+    curl -sS --max-time 3 'http://localhost:3473/' --data-binary 'SELECT 1' 2>&1
+    echo "=== ss listeners ==="
+    sudo ss -lntp 2>&1 | head -20
+} >&2
+exit 1

firebolt-parquet-partitioned/stop

@@ -0,0 +1,8 @@
+#!/bin/bash
+set -e
+
+# Plain stop — leave the container in place so its bind-mounted
+# fb-volume keeps the loaded database for the next ./start. The
+# container is removed and the volume re-initialised only on
+# explicit re-provision.
+sudo docker container stop firebolt-core >/dev/null 2>&1 || true