diff --git a/config/runtime.exs b/config/runtime.exs
index b1d73c3a..e179efeb 100644
--- a/config/runtime.exs
+++ b/config/runtime.exs
@@ -815,6 +815,24 @@ if channels = System.get_env("GEN_RPC_SCATTER_CHANNELS") do
     channels: Smolquery.RuntimeConfig.positive_integer!("GEN_RPC_SCATTER_CHANNELS", channels)
 end
 
+# T-614: how long a socket to a peer node may outlive the peer.
+if ms = System.get_env("SMOLQUERY_PEER_TCP_USER_TIMEOUT_MS") do
+  config :smolquery, Smolquery.PeerSocket,
+    user_timeout_ms:
+      Smolquery.RuntimeConfig.positive_integer!("SMOLQUERY_PEER_TCP_USER_TIMEOUT_MS", ms)
+end
+
+if seconds = System.get_env("SMOLQUERY_PEER_TCP_KEEPALIVE_IDLE_S") do
+  config :smolquery, Smolquery.PeerSocket,
+    keepalive_idle_s:
+      Smolquery.RuntimeConfig.integer_in_range!(
+        "SMOLQUERY_PEER_TCP_KEEPALIVE_IDLE_S",
+        seconds,
+        1,
+        32_767
+      )
+end
+
 if catalog_database_url = System.get_env("CATALOG_DATABASE_URL") do
   db = Smolquery.DatabaseUrl.parse!(catalog_database_url)
 
diff --git a/docs/configuration.md b/docs/configuration.md
index 8eb3720a..ea05d31d 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -169,6 +169,8 @@ version's driver exists.
 | `GEN_RPC_SCATTER_CHANNELS` | How many `{:scatter, _}` gen_rpc connections a node opens to each peer for distributed-query partials (`4`). A job hashes to one connection; concurrent jobs spread over the pool |
 | `GEN_RPC_TLS` | `true` switches buffer/query inter-node traffic to mutual TLS (Transport Layer Security) (`false`). Verification is chain-only against the cluster CA (certificate authority); the emqx gen_rpc fork does no hostname or CN (common name) check. The CA is thus the trust boundary. Certificate files are per node (`GEN_RPC_TLS_DIR`, default `/etc/smolquery/gen-rpc-tls`; `POD_NAME` names the file). Any CA-signed certificate authenticates to any peer. A leaked node certificate thus means a rotation of the CA, not just of that node |
 | `GEN_RPC_SSL_PORT` | The gen_rpc TLS port (`5870`) |
+| `SMOLQUERY_PEER_TCP_USER_TIMEOUT_MS` | How long a socket to a peer node may go without an acknowledgement before the kernel closes it (`15000`): sent data in flight, or keepalive probes on an idle socket. Applies to `HotClient`'s manifest reads (T-614). Linux only |
+| `SMOLQUERY_PEER_TCP_KEEPALIVE_IDLE_S` | How long a socket to a peer node may sit idle before the kernel starts keepalive probes (`5`, at most `32767`). Probes repeat every 5 s; `SMOLQUERY_PEER_TCP_USER_TIMEOUT_MS`, not a probe count, decides when an unanswered idle socket closes. Linux only for the timers; elsewhere keepalive uses the OS defaults |
 | `DIST_TLS` | `true` runs Erlang distribution (cluster membership only) over TLS with the same certificates (`false`). Set it in `rel/env.sh.eex`, not `config/runtime.exs`. Distribution starts before the release's Elixir config does |
 | `POD_NAME` / `POD_NAMESPACE` | When set (a Kubernetes Downward API convention), `rel/env.sh.eex` derives `RELEASE_NODE` from the pod's stable headless-service DNS name. That is the same name a peer needs to reach this node |
 | `HEADLESS_SERVICE` | The headless-service name in that derived node name (`smolquery-headless`) |
diff --git a/docs/deployment.md b/docs/deployment.md
index 591e91cc..7b15026a 100644
--- a/docs/deployment.md
+++ b/docs/deployment.md
@@ -97,10 +97,42 @@ and `SMOLQUERY_DISTRIBUTED_WORKER_THREADS`. A scattered query's declared
 budget on one node is the worker count `×` the worker limit, on top of the
 job engine's own `job_memory_limit`.
 
+## When a peer dies without closing its sockets
+
+A pod that is killed hard closes none of its sockets. Its replacement has the
+same name and a new IP. Linux keeps a socket to the old IP until it gives up
+retransmitting: `tcp_retries2` = 15, about 15.4 minutes. On 2026-10-02 four
+hard-killed pods were replaced in 13 s, and the cluster answered 5xx for 16
+minutes.
+
+Each connection between nodes now has its own bound:
+
+| connection | what bounds it | setting |
+|---|---|---|
+| gen_rpc, buffer writes and replication (`:control`, `{:bulk, _}`) and scatter partials (`{:scatter, _}`) | Every client to a node is killed when distribution reports it down (T-613). A call that times out probes its channel and kills the client if the probe is unanswered (T-613). | none for the drop; the probe waits 10 s |
+| gen_rpc sockets this node accepts | gen_rpc's own kernel keepalive: idle 5 s, interval 5 s, 2 probes | `config :gen_rpc, socket_keepalive_*` |
+| gen_rpc sockets this node dials | `SO_KEEPALIVE` is on, with Linux's default timers: first probe after 2 h. gen_rpc 3.6.1 does not let an application set the timers or `TCP_USER_TIMEOUT` on a client socket. The first row covers it. The pod sysctls `net.ipv4.tcp_keepalive_time`, `_intvl` and `_probes` (Kubernetes safe sysctls) would shorten those timers for every socket in the pod. | none in smolquery |
+| `HotClient` HTTP: the storage sealer and the query planner reading a buffer node's manifest | `TCP_USER_TIMEOUT` closes a socket to a dead peer after 15 s, and the pool drops it (T-614). Keepalive probes an idle socket every 5 s after 5 s idle, so an idle socket is checked against the same 15 s; on Linux the user timeout, not the probe count, decides when keepalive gives up. | `SMOLQUERY_PEER_TCP_USER_TIMEOUT_MS`, `SMOLQUERY_PEER_TCP_KEEPALIVE_IDLE_S`, `config :smolquery, Smolquery.PeerSocket` |
+| Erlang distribution | The net tick: a silent node is down after `net_ticktime`, 60 s by default. On 2026-10-02 it reconnected in seconds. | `-kernel net_ticktime` |
+| DuckDB `httpfs` reading hot segments from a buffer node | DuckDB's own `http_timeout`. Not bounded by smolquery. | none |
+
+The keepalive timers and `TCP_USER_TIMEOUT` are Linux socket options. On
+another OS only keepalive is switched on, with that OS's timers.
+
+Neither bound cuts a slow peer short. A slow peer's kernel still
+acknowledges data and keepalive probes; the bounds act only on a peer that
+answers nothing at the TCP level.
+
 ## Upgrade notes
 
 One note per release, newest first.
 
+### 0.22.0: sockets to a replaced pod close in seconds (T-613, T-614)
+
+- **gen_rpc:** a node's clients to a peer are killed when distribution reports the peer down, and a channel that times out is probed and redialed if it is dead (T-613). Nothing to configure.
+- **Buffer manifest reads (`HotClient`)** close a socket to a dead peer after 15 s: a 15 s `TCP_USER_TIMEOUT`, with keepalive probing an idle socket every 5 s after 5 s idle. Before, a pooled connection to a replaced buffer pod stayed in the pool and failed one seal per connection, each after the full 30 s receive timeout, about once a minute. Tune with `SMOLQUERY_PEER_TCP_USER_TIMEOUT_MS` and `SMOLQUERY_PEER_TCP_KEEPALIVE_IDLE_S`.
+- See [When a peer dies without closing its sockets](#when-a-peer-dies-without-closing-its-sockets) for every connection's bound.
+
 ### 0.21.0: federated DuckLake connections (T-610)
 
 You can register another DuckLake as a connection and query it: `"kind": "ducklake"` on `POST /v1/connections`, or the kind selector on the connections page. It takes the lake's Postgres metadata database, its data path and optional S3 credentials, sealed like a password. The lake attaches read-only at its data path, and a query reads it as `name.schema.table`, alone or joined with smolquery tables. See [api.md](api.md).
diff --git a/lib/smolquery/buffer_service/hot_client.ex b/lib/smolquery/buffer_service/hot_client.ex
index 21672552..c7a5ac02 100644
--- a/lib/smolquery/buffer_service/hot_client.ex
+++ b/lib/smolquery/buffer_service/hot_client.ex
@@ -14,8 +14,19 @@ defmodule Smolquery.BufferService.HotClient do
   `HotServer` listens is the caller's configuration (`buffer_base_url` on the
   storage and query runtimes) — honest for a single node; a cluster resolves
   it from the ownership ring instead, which arrives with Milestone 8.
+
+  ## Connections to a replaced pod
+
+  Requests go through a pooled connection with `Smolquery.PeerSocket`'s
+  keepalive timers and `TCP_USER_TIMEOUT` (T-614). A buffer pod killed
+  without closing its sockets leaves pooled connections to its old IP.
+  Without those bounds each one stayed in the pool until a request drew it
+  and waited out the whole receive timeout, one failed seal per stale
+  connection. With them, the kernel closes a dead socket after 15 s, idle or
+  in flight, and the pool drops it.
   """
 
+  alias Smolquery.PeerSocket
   alias Smolquery.Segments.Store
 
   @default_timeout_ms 30_000
@@ -128,7 +139,14 @@ defmodule Smolquery.BufferService.HotClient do
 
   defp fetch(url, scope, timeout_ms) do
     headers = [{Smolquery.InternalSecret.header(), Smolquery.InternalSecret.value()}]
-    common = [url: url, headers: headers, receive_timeout: timeout_ms, retry: false]
+
+    common = [
+      url: url,
+      headers: headers,
+      receive_timeout: timeout_ms,
+      retry: false,
+      connect_options: [transport_opts: PeerSocket.tcp_options()]
+    ]
 
     request =
       if scope, do: [method: :post, json: scope] ++ common, else: [method: :get] ++ common
diff --git a/lib/smolquery/peer_socket.ex b/lib/smolquery/peer_socket.ex
new file mode 100644
index 00000000..a6353ebb
--- /dev/null
+++ b/lib/smolquery/peer_socket.ex
@@ -0,0 +1,81 @@
+defmodule Smolquery.PeerSocket do
+  @moduledoc """
+  TCP options that bound how long a socket to a peer node can outlive the peer
+  (T-614).
+
+  A pod killed without closing its sockets leaves every connection to it
+  pointed at an IP that no longer answers. Linux keeps such a socket until it
+  gives up retransmitting, `tcp_retries2` = 15, about 15.4 minutes. Finch
+  turns `SO_KEEPALIVE` on for every pooled socket, but with Linux's default
+  timers an idle socket is first probed after two hours. A pooled HTTP
+  connection to a dead peer therefore stays in its pool, and each request
+  that draws it waits out its whole receive timeout before the pool lets it
+  go. On the sandbox on 2026-10-02 that was a failed seal about once a minute
+  after a buffer pod was replaced, one per stale connection.
+
+  `tcp_options/0` sets one bound, `user_timeout_ms` (15,000), through two
+  socket options:
+
+    * **`TCP_USER_TIMEOUT`** closes a socket whose sent data has gone
+      unacknowledged that long. A request already in flight on a dead
+      socket fails then, not after its receive timeout.
+    * **Keepalive** probes an idle socket after `keepalive_idle_s` (5) and
+      every `keepalive_interval_s` (5) after that. On Linux the user timeout
+      also governs keepalive: the kernel closes an idle socket at the first
+      unanswered probe at or after `user_timeout_ms`, and the probe count
+      plays no part. At the defaults that is 15 s. A closed idle socket
+      leaves its pool before a request can draw it.
+
+  So `user_timeout_ms` is the bound for both cases, and the keepalive timers
+  only set how often an idle socket is checked against it.
+
+  Both act only on a peer that has stopped answering at the TCP level. A slow
+  peer still ACKs data and probes, so neither cuts a slow response short.
+
+  The timers are Linux socket options set through `{:raw, ...}`. Elsewhere only
+  `keepalive: true` is set, with the operating system's own timers.
+
+      config :smolquery, Smolquery.PeerSocket,
+        user_timeout_ms: 15_000,
+        keepalive_idle_s: 5,
+        keepalive_interval_s: 5
+  """
+
+  @ipproto_tcp 6
+  @tcp_keepidle 4
+  @tcp_keepintvl 5
+  @tcp_user_timeout 18
+
+  @defaults [user_timeout_ms: 15_000, keepalive_idle_s: 5, keepalive_interval_s: 5]
+
+  @doc """
+  The `:gen_tcp` options for a socket to a peer node, from this node's
+  configuration.
+  """
+  @spec tcp_options() :: keyword()
+  def tcp_options, do: tcp_options(config(), :os.type())
+
+  @doc """
+  The `:gen_tcp` options for `config` on the operating system `os_type`, as
+  `:os.type/0` names it.
+  """
+  @spec tcp_options(keyword(), {atom(), atom()}) :: keyword()
+  def tcp_options(config, {:unix, :linux}) do
+    [
+      keepalive: true,
+      raw: {@ipproto_tcp, @tcp_keepidle, int(config[:keepalive_idle_s])},
+      raw: {@ipproto_tcp, @tcp_keepintvl, int(config[:keepalive_interval_s])},
+      raw: {@ipproto_tcp, @tcp_user_timeout, int(config[:user_timeout_ms])}
+    ]
+  end
+
+  def tcp_options(_config, _os_type), do: [keepalive: true]
+
+  @doc """
+  This node's settings, each defaulted.
+  """
+  @spec config() :: keyword()
+  def config, do: Keyword.merge(@defaults, Application.get_env(:smolquery, __MODULE__, []))
+
+  defp int(value), do: <<value::32-native>>
+end
diff --git a/test/smolquery/buffer_service/hot_client_test.exs b/test/smolquery/buffer_service/hot_client_test.exs
index fefd2419..7d12b4ad 100644
--- a/test/smolquery/buffer_service/hot_client_test.exs
+++ b/test/smolquery/buffer_service/hot_client_test.exs
@@ -10,6 +10,7 @@ defmodule Smolquery.BufferService.HotClientTest do
   @moduletag :tmp_dir
 
   @table {"analytics", "events"}
+  @linux match?({:unix, :linux}, :os.type())
 
   defp batch(range) do
     %{schema: Schema.new!([{"id", :int64}]), rows: for(i <- range, do: %{"id" => i})}
@@ -163,4 +164,36 @@ defmodule Smolquery.BufferService.HotClientTest do
                {:error, {:invalid_identifier, "../etc"}}
     end
   end
+
+  describe "connections (T-614)" do
+    @tag skip: not @linux and "Linux socket options"
+    test "the pooled socket to a buffer node carries PeerSocket's bounds", context do
+      buffer = start_buffer(context)
+      base_url = HotServer.base_url(buffer)
+      port = base_url |> URI.parse() |> Map.fetch!(:port)
+
+      assert {:ok, []} = HotClient.manifest(base_url, @table)
+
+      [socket | _rest] = sockets_to(port)
+
+      {:ok, options} =
+        :inet.getopts(socket, [:keepalive] ++ for(o <- [4, 5, 18], do: {:raw, 6, o, 4}))
+
+      assert {:keepalive, true} in options
+
+      assert for(
+               {:raw, 6, option, <<value::32-native>>} <- options,
+               into: %{},
+               do: {option, value}
+             ) ==
+               %{4 => 5, 5 => 5, 18 => 15_000}
+    end
+  end
+
+  defp sockets_to(port) do
+    for socket <- Port.list(),
+        Port.info(socket, :name) == {:name, ~c"tcp_inet"},
+        match?({:ok, {_address, ^port}}, :inet.peername(socket)),
+        do: socket
+  end
 end
diff --git a/test/smolquery/peer_socket_test.exs b/test/smolquery/peer_socket_test.exs
new file mode 100644
index 00000000..317d3e85
--- /dev/null
+++ b/test/smolquery/peer_socket_test.exs
@@ -0,0 +1,55 @@
+defmodule Smolquery.PeerSocketTest do
+  use ExUnit.Case, async: false
+
+  alias Smolquery.PeerSocket
+
+  @linux match?({:unix, :linux}, :os.type())
+
+  defp read_back(options) do
+    for {:raw, 6, option, <<value::32-native>>} <- options, into: %{}, do: {option, value}
+  end
+
+  defp decoded(options) do
+    for {:raw, {6, option, <<value::32-native>>}} <- options, into: %{}, do: {option, value}
+  end
+
+  test "on Linux sets keepalive timers and TCP_USER_TIMEOUT" do
+    options = PeerSocket.tcp_options(PeerSocket.config(), {:unix, :linux})
+
+    assert options[:keepalive] == true
+    assert decoded(options) == %{4 => 5, 5 => 5, 18 => 15_000}
+  end
+
+  test "elsewhere sets keepalive only" do
+    assert PeerSocket.tcp_options(PeerSocket.config(), {:unix, :darwin}) == [keepalive: true]
+  end
+
+  test "config overrides each default" do
+    previous = Application.fetch_env(:smolquery, PeerSocket)
+    Application.put_env(:smolquery, PeerSocket, user_timeout_ms: 7_000, keepalive_idle_s: 3)
+
+    on_exit(fn ->
+      case previous do
+        {:ok, value} -> Application.put_env(:smolquery, PeerSocket, value)
+        :error -> Application.delete_env(:smolquery, PeerSocket)
+      end
+    end)
+
+    assert PeerSocket.config()[:keepalive_interval_s] == 5
+    assert decoded(PeerSocket.tcp_options(PeerSocket.config(), {:unix, :linux}))[18] == 7_000
+    assert decoded(PeerSocket.tcp_options(PeerSocket.config(), {:unix, :linux}))[4] == 3
+  end
+
+  @tag skip: not @linux and "Linux socket options"
+  test "a socket opened with tcp_options/0 carries them" do
+    {:ok, listener} = :gen_tcp.listen(0, [:binary, active: false])
+    {:ok, port} = :inet.port(listener)
+    {:ok, socket} = :gen_tcp.connect(~c"127.0.0.1", port, [:binary] ++ PeerSocket.tcp_options())
+
+    {:ok, options} =
+      :inet.getopts(socket, [:keepalive] ++ for(o <- [4, 5, 18], do: {:raw, 6, o, 4}))
+
+    assert {:keepalive, true} in options
+    assert read_back(options) == %{4 => 5, 5 => 5, 18 => 15_000}
+  end
+end