diff --git a/apisix/plugins/prometheus/exporter.lua b/apisix/plugins/prometheus/exporter.lua
index fbb600e18687..0fdfbb3049ca 100644
--- a/apisix/plugins/prometheus/exporter.lua
+++ b/apisix/plugins/prometheus/exporter.lua
@@ -89,6 +89,21 @@ end
 -- Default refresh interval
 local DEFAULT_REFRESH_INTERVAL = 15
 
+-- Default interval for reclaiming expired entries from the metrics shared dict
+local DEFAULT_FLUSH_EXPIRED_INTERVAL = 60
+
+-- Entries a single flush_expired() call may reclaim. The call holds the shared
+-- dict lock until it returns, so this is what bounds how long the workers can
+-- be kept waiting on that lock: measured at ~1ms per 10000 entries.
+local FLUSH_EXPIRED_BATCH = 10000
+
+-- Upper bound on the batches run by one tick, so the loop always finishes well
+-- within one interval. Whatever is left over is picked up by the next tick.
+local FLUSH_EXPIRED_MAX_BATCHES = 30
+
+-- Pause between batches, so the lock is not taken back to back
+local FLUSH_EXPIRED_BATCH_DELAY = 1
+
 local CACHED_METRICS_KEY = "cached_metrics_text"
 
 local metrics = {}
@@ -99,6 +114,10 @@ local exporter_timer_running = false
 
 local exporter_timer_created = false
 
+local flush_expired_timer_running = false
+
+local flush_expired_timer_created = false
+
 local function gen_arr(...)
     clear_tab(inner_tab_arr)
     for i = 1, select('#', ...) do
@@ -1201,6 +1220,105 @@ local function exporter_timer(premature, yieldable, cache_exptime)
 end
 
 
+-- Expired entries in the metrics shared dict are only logically dead: every
+-- dict API reports them as missing, but their slab pages stay allocated until
+-- something reclaims them. The passive per-write expiry scan cannot do it,
+-- because it stops at the first non-expired entry at the LRU tail and a
+-- permanent entry (the error metric, or any metric registered without an
+-- expire) inevitably ends up sitting there. Left alone, the dict fills up with
+-- dead entries and starts evicting live ones (apache/apisix#13658). The metrics
+-- library reclaims them from every worker once an hour with an unbounded
+-- flush_expired() call, which walks the whole LRU queue with the dict lock
+-- held. Draining them here instead -- in the privileged agent, in bounded
+-- batches -- keeps that backlog, and therefore the lock hold of any single
+-- call, small.
+local function flush_expired_metrics()
+    local dict = ngx.shared["prometheus-metrics"]
+    if not dict then
+        return 0
+    end
+
+    local total = 0
+    for _ = 1, FLUSH_EXPIRED_MAX_BATCHES do
+        -- a return value below the batch size means this call has already
+        -- walked the whole LRU queue, so there is nothing left to reclaim
+        local freed = dict:flush_expired(FLUSH_EXPIRED_BATCH)
+        total = total + freed
+        if freed < FLUSH_EXPIRED_BATCH then
+            break
+        end
+
+        ngx.sleep(FLUSH_EXPIRED_BATCH_DELAY)
+    end
+
+    return total
+end
+_M.flush_expired_metrics = flush_expired_metrics
+
+
+local function flush_expired_timer(premature)
+    if premature then
+        return
+    end
+
+    if flush_expired_timer_running then
+        core.log.warn("Previous metrics flush still running, skipping")
+        return
+    end
+
+    flush_expired_timer_running = true
+
+    local ok, err = pcall(flush_expired_metrics)
+    if not ok then
+        core.log.error("Failed to flush expired metrics: ", err)
+    end
+
+    flush_expired_timer_running = false
+end
+
+
+-- Nothing in the dict ever becomes expired unless at least one metric is
+-- registered with an expire, so the timer is only worth running then.
+local function metrics_expire_enabled(attr)
+    local metrics_attr = attr and attr.metrics
+    if type(metrics_attr) ~= "table" then
+        return false
+    end
+
+    for _, conf in pairs(metrics_attr) do
+        local expire = type(conf) == "table" and tonumber(conf.expire)
+        if expire and expire > 0 then
+            return true
+        end
+    end
+
+    return false
+end
+
+
+local function init_flush_expired_timer(attr)
+    if flush_expired_timer_created or not metrics_expire_enabled(attr) then
+        return
+    end
+
+    local interval = DEFAULT_FLUSH_EXPIRED_INTERVAL
+    if attr and attr.flush_expired_interval then
+        interval = attr.flush_expired_interval
+    end
+
+    if interval <= 0 then
+        return
+    end
+
+    local ok, err = ngx_timer_every(interval, flush_expired_timer)
+    if ok then
+        flush_expired_timer_created = true
+    else
+        core.log.error("Failed to start the metrics flush timer: ", err)
+    end
+end
+
+
 local function init_exporter_timer()
     if process.type() ~= "privileged agent" then
         return
@@ -1214,6 +1332,8 @@ local function init_exporter_timer()
 
     local cache_exptime = refresh_interval * 2
 
+    init_flush_expired_timer(attr)
+
     exporter_timer(false, false, cache_exptime)
 
     if exporter_timer_created then
diff --git a/conf/config.yaml.example b/conf/config.yaml.example
index 6a1361aec496..bb93121fdc38 100644
--- a/conf/config.yaml.example
+++ b/conf/config.yaml.example
@@ -724,6 +724,9 @@ plugin_attr:          # Plugin attributes
       ip: 127.0.0.1                         # Set the IP.
       port: 9091                            # Set the port.
     refresh_interval: 15                    # Set the interval for refreshing cached metric data. unit: second.
+    flush_expired_interval: 60              # Set the interval for reclaiming expired entries from the
+                                            # `prometheus-metrics` shared dict. Only used when a metric
+                                            # below is configured with `expire`. unit: second.
     # metrics:    # Create extra labels from nginx variables: https://nginx.org/en/docs/varindex.html
     #  http_status:
     #    expire: 0 # The expiration time after which metrics are removed. unit: second.
diff --git a/t/plugin/prometheus-metric-expire.t b/t/plugin/prometheus-metric-expire.t
index caad85ea04a5..dd30201ff390 100644
--- a/t/plugin/prometheus-metric-expire.t
+++ b/t/plugin/prometheus-metric-expire.t
@@ -130,3 +130,34 @@ plugin_attr:
 GET /t
 --- response_body
 passed
+
+
+
+=== TEST 2: expired entries are reclaimed in bounded batches
+--- config
+    location /t {
+        content_by_lua_block {
+            local exporter = require("apisix.plugins.prometheus.exporter")
+            local dict = ngx.shared["prometheus-metrics"]
+
+            -- a permanent entry kept at the LRU tail stops the passive
+            -- per-write expiry scan, which is what lets expired entries pile
+            -- up in the first place
+            dict:set("flush_expired_tail", 1)
+            for i = 1, 20000 do
+                dict:set("flush_expired_" .. i, "v", 0.1)
+            end
+            ngx.sleep(0.2)
+
+            -- more than one batch, so the loop has to run again after the
+            -- delay instead of stopping at the first batch
+            local freed = exporter.flush_expired_metrics()
+            ngx.say("reclaimed all: ", freed >= 20000)
+            ngx.say("left over: ", exporter.flush_expired_metrics())
+            ngx.say("tail kept: ", dict:get("flush_expired_tail"))
+        }
+    }
+--- response_body
+reclaimed all: true
+left over: 0
+tail kept: 1