diff --git a/src/api.rs b/src/api.rs
index 9949dc27..d88606b2 100644
--- a/src/api.rs
+++ b/src/api.rs
@@ -1327,8 +1327,8 @@ fn scratch_path(app: &App, kind: &str) -> String {
 /// The data page's figures.
 ///
 /// Off the runtime, like the three routes below: `stats` counts every row of
-/// `metric` and `ping_record` -- both WITHOUT ROWID, so each count is a full
-/// index scan -- holding the connection the agents report through throughout. At
+/// both tiers of history -- all WITHOUT ROWID, so each count is a full index
+/// scan -- holding the connection the agents report through throughout. At
 /// 2.2M rows that is 127 ms during which the public page and every agent report
 /// also wait, growing with `retention_days`.
 pub async fn db_stats(_: Admin, State(app): State<Shared>) -> Response {
diff --git a/src/db.rs b/src/db.rs
index 3c899e63..86422398 100644
--- a/src/db.rs
+++ b/src/db.rs
@@ -1520,7 +1520,11 @@ impl Db {
     /// `minutes` is how many minute rows the bucket holds, against the
     /// `step / 60` it spans: a node offline for part of a bucket has its means
     /// taken over the minutes it reported, and a caller integrating the rates
-    /// or showing availability needs the difference.
+    /// or showing availability needs the difference. That holds while the
+    /// agent reports at least once a minute. A row is written only in a minute
+    /// a report arrives, so past a 60-second `--interval` each row covers one
+    /// interval and `minutes` falls in proportion: a fifth of the full count at
+    /// 300 seconds.
     ///
     /// An hourly window reads the folded hours before the watermark and the
     /// minute rows after it, each hour weighted by the minutes it holds, so a
@@ -1691,11 +1695,18 @@ impl Db {
     /// `delete_node` clears the two probe tables, and `insert_pings` writes only
     /// for probes assigned to an existing node.
     ///
-    /// At most a day of one node's rows per statement, with the lock the agents
-    /// write through released in between. An hourly pass deletes an hour and is
-    /// one statement per table; the first pass after an upgrade deletes
-    /// everything past the week, which in one statement would hold the lock for
-    /// 37.8 s at 90 days of 100 nodes, and a node at a time 0.9 s each.
+    /// Minute rows are deleted at most one node's day per statement, with the
+    /// lock the agents write through released in between. An hourly pass
+    /// deletes an hour and is one statement per table; the first pass after an
+    /// upgrade deletes everything past the week, which in one statement would
+    /// hold the lock for 37.8 s at 90 days of 100 nodes, and a node at a time
+    /// 0.9 s each.
+    ///
+    /// Hourly rows are deleted a node at a time, one statement per table:
+    /// lowering the window from a year to a month deletes 8,040 and 32,160 rows
+    /// of a node with four probes, 24 ms each. A day per statement would spend
+    /// 0.96 s per node on the lookups and pauses in between, 4.8 minutes at 300
+    /// nodes ahead of the `VACUUM` that usually follows.
     ///
     /// The watermark is read with each statement: a restore can replace it,
     /// and the minute rows it guards, between two of them.
@@ -1729,7 +1740,7 @@ impl Db {
                     let Some(first) = first.filter(|&ts| ts < before) else { break };
                     pruned += conn
                         .prepare_cached(&format!("DELETE FROM {table} WHERE node_id=?1 AND ts<?2"))?
-                        .execute(params![id, before.min(first + 86_400)])?;
+                        .execute(params![id, if guarded { before.min(first + 86_400) } else { before }])?;
                     drop(conn);
                     let_waiters_in();
                 }
diff --git a/web-admin/src/components/Admin.tsx b/web-admin/src/components/Admin.tsx
index c27d3072..3e196735 100644
--- a/web-admin/src/components/Admin.tsx
+++ b/web-admin/src/components/Admin.tsx
@@ -2883,11 +2883,13 @@ type DbInfo = {
   rows: Record<string, number>
 }
 
-// The only two tables whose row count indicates anything about size. Every other
-// holds one row per node or per key.
+// The only tables whose row count indicates anything about size, each kind of
+// history in both tiers. Every other holds one row per node or per key.
 const DB_ROWS: [string, string][] = [
   ["metric", "历史明细"],
+  ["metric_hour", "历史小时汇总"],
   ["ping_record", "延迟记录"],
+  ["ping_hour", "延迟小时汇总"],
 ]
 
 function Data() {
@@ -2941,8 +2943,8 @@ function Data() {
   }
 
   if (!info) return null
-  const stat = (label: string, value: string) => (
-    <div key={label}>
+  const stat = (label: string, value: string, className = "") => (
+    <div key={label} className={className}>
       <div className="text-xs text-muted-foreground">{label}</div>
       <div className="tnum mt-0.5 text-sm">{value}</div>
     </div>
@@ -2952,10 +2954,13 @@ function Data() {
     <div className="space-y-4">
       <Card className="gap-4 p-5">
         <h3 className="text-sm font-medium">数据库</h3>
-        <div className="grid grid-cols-2 gap-4 sm:grid-cols-4">
+        {/* Five columns: the file and the window on one row, the four row counts
+            on the next. On two columns the free space takes a row of its own, so
+            the window and each kind's two tiers still pair up. */}
+        <div className="grid grid-cols-2 gap-4 sm:grid-cols-5">
           {stat("文件大小", bytes(info.size))}
           {stat("预写日志", bytes(info.wal))}
-          {stat("可回收空间", bytes(info.free))}
+          {stat("可回收空间", bytes(info.free), "col-span-2 sm:col-span-1")}
           {stat("保留天数", `${info.retention} 天`)}
           {/* 和保留天数并排：跨度小于保留期是还没攒够，大于保留期就是每小时
               那次 prune 没在跑。 */}