diff --git a/modules/monitoring.nix b/modules/monitoring.nix
index 7865202..2a7f7b5 100644
--- a/modules/monitoring.nix
+++ b/modules/monitoring.nix
@@ -60,21 +60,6 @@
       fi
     done
   '';
-
-  # Physical pool IOPS/bandwidth straight from the vdev layer. The telegraf
-  # `inputs.zfs` poolMetrics emits per-objset *logical* counters, which include
-  # ARC cache hits — so RAM-served reads show up as huge "pool IOPS" with zero
-  # actual disk activity (this misled a 2026-05-30 mountd investigation). This
-  # collector reports what `zpool iostat` reports: real disk ops. The second
-  # block of `iostat 1 2` is the live 1s sample; the operations columns are
-  # already per-second rates, so these are gauges (no derivative in Grafana).
-  zpoolIostatScript = pkgs.writeShellScript "zpool-iostat" ''
-    npools=$(${pkgs.zfs}/bin/zpool list -H -o name | ${pkgs.coreutils}/bin/wc -l)
-    ${pkgs.zfs}/bin/zpool iostat -Hp 1 2 | ${pkgs.coreutils}/bin/tail -n "$npools" \
-      | while read -r name alloc free rops wops rbytes wbytes; do
-        echo "zpool_io,pool=$name read_ops=''${rops}i,write_ops=''${wops}i,read_bytes=''${rbytes}i,write_bytes=''${wbytes}i"
-      done
-  '';
 in {
   options.skyworks.monitoring = {
     enable = lib.mkEnableOption "Telegraf monitoring to InfluxDB";
@@ -166,14 +151,6 @@
               timeout = "60s";
               data_format = "influx";
             }
-            {
-              # Physical pool IOPS (measurement `zpool_io`) — the honest version
-              # of the ARC-inclusive `zfs` poolMetrics counters.
-              commands = [ "${zpoolIostatScript}" ];
-              interval = "10s";
-              timeout = "10s";
-              data_format = "influx";
-            }
           ];
         };
       };
@@ -279,6 +256,31 @@
           } > "$TMP_ZFS"
           chmod 0644 "$TMP_ZFS"
           mv -f "$TMP_ZFS" "$OUT_DIR/zpool_health.prom"
+
+          # ZFS pool -> block-device mapping. Lets Grafana compute *physical*
+          # per-pool IOPS by joining node_disk_*_completed_total on `device`
+          # (node_zfs_zpool_dataset_* is LOGICAL — it counts ARC cache hits, so
+          # RAM-served reads look like huge phantom pool IOPS). Re-read from live
+          # `zpool status` each run, so kernel device renames across reboots
+          # can't break the dashboard. Partition suffix is stripped to match
+          # node-exporter's whole-disk `device` label (sdb1->sdb, nvme0n1p1->nvme0n1).
+          TMP_MEM=$(mktemp "$OUT_DIR/zpool_members.prom.XXXXXX")
+          {
+            echo "# HELP skyw_zpool_member 1 if the block device is a member of the pool"
+            echo "# TYPE skyw_zpool_member gauge"
+            ${pkgs.zfs}/bin/zpool status -LP 2>/dev/null | ${pkgs.gawk}/bin/awk '
+              /^[[:space:]]*pool:/ { pool = $2; next }
+              /\/dev\// {
+                for (i = 1; i <= NF; i++) if ($i ~ /^\/dev\//) {
+                  dev = $i; sub(/^\/dev\//, "", dev)
+                  if (dev ~ /^nvme/) sub(/p[0-9]+$/, "", dev); else sub(/[0-9]+$/, "", dev)
+                  line = "skyw_zpool_member{pool=\"" pool "\",device=\"" dev "\"} 1"
+                  if (!(line in seen)) { seen[line] = 1; print line }
+                }
+              }'
+          } > "$TMP_MEM"
+          chmod 0644 "$TMP_MEM"
+          mv -f "$TMP_MEM" "$OUT_DIR/zpool_members.prom"
         '';
       };
     };