SCH-7988 raise the server-mode reservation TTL cap to 24 hours

README.md

@@ -707,7 +707,7 @@ All fields live on `CreditLeaseConfig`. Durations are seconds. Everything below
 | Option | Type | Default | Description |
 |---|---|---|---|
 | `mode` | `str` | `"auto"` | `client`, `server`, or `auto` (client when DataStream is running). |
-| `default_reservation_ttl` | `float` | 60 | How long an unsettled hold survives. Capped at one hour in server mode. |
+| `default_reservation_ttl` | `float` | 60 | How long an unsettled hold survives. Capped at 24 hours in server mode. |
 | `default_lease_duration` | `float` | 300 | Lease lifetime requested at acquire and extend. |
 | `default_lease_size` | `float` | 10000 | Credits requested per acquire, and the minimum extend tranche. |
 | `low_water_mark` | `float` | 0.25 | Extend in the background when the lease balance dips below this fraction. |

src/schematic/client.py

@@ -86,7 +86,7 @@
 
 # The longest hold the server will take. A longer configured TTL is clamped to
 # it, rather than sent and rejected on every check.
-MAX_RESERVATION_TTL = 3600.0
+MAX_RESERVATION_TTL = 86400.0
 # Room for the two clocks to disagree. expires_at is computed here and measured
 # against the server's own clock, so a TTL sitting exactly on the cap is
 # rejected whenever this process runs even slightly ahead.
@@ -159,7 +159,7 @@ class CreditLeaseConfig:
     mode: CreditLeaseMode = "auto"
     # How long a hold survives unsettled. Size it above the longest expected
     # gap between check() and track_with_reservation(). Anything above the
-    # server's one hour cap is clamped, in server mode. Client mode keeps it:
+    # server's 24 hour cap is clamped, in server mode. Client mode keeps it:
     # the TTL only drives the local sweeper there.
     default_reservation_ttl: float = DEFAULT_RESERVATION_TTL
     # Lease lifetime requested at acquire and extend. Default 5 minutes.
@@ -408,7 +408,7 @@ def _resolve_reservation_ttl(logger: logging.Logger, credit_leases: Optional[Cre
 
     Client mode is exempt: its TTL never reaches the server, it only tells the
     local sweeper when to refund an unsettled hold, and the caller may well
-    want one that outlives an hour.
+    want one that outlives a day.
     """
     if credit_leases is None:
         return DEFAULT_RESERVATION_TTL
@@ -418,7 +418,7 @@ def _resolve_reservation_ttl(logger: logging.Logger, credit_leases: Optional[Cre
     effective = MAX_RESERVATION_TTL - RESERVATION_TTL_SKEW_ALLOWANCE
     if ttl > effective:
         logger.warning(
-            f"credit_leases.default_reservation_ttl of {ttl}s is above the server's one hour cap; "
+            f"credit_leases.default_reservation_ttl of {ttl}s is above the server's 24 hour cap; "
             f"server-mode holds will be clamped to {effective}s"
         )
         return effective

tests/custom/test_client.py

@@ -1842,13 +1842,13 @@ def test_an_integral_float_usage_reaches_the_preflight_unchanged(self):
         )
 
     def test_a_reservation_ttl_above_the_cap_is_clamped(self):
-        client = self._client(credit_leases=CreditLeaseConfig(default_reservation_ttl=7200.0))
+        client = self._client(credit_leases=CreditLeaseConfig(default_reservation_ttl=90000.0))
         try:
             # Short of the cap by the skew allowance, so a client running
             # slightly fast still asks for something the server accepts.
             self.assertEqual(client._reservation_ttl, MAX_RESERVATION_TTL - RESERVATION_TTL_SKEW_ALLOWANCE)
             warning = " ".join(str(call.args[0]) for call in client.logger.warning.call_args_list)
-            self.assertIn("one hour cap", warning)
+            self.assertIn("24 hour cap", warning)
             self.assertIn(
                 f"server-mode holds will be clamped to {MAX_RESERVATION_TTL - RESERVATION_TTL_SKEW_ALLOWANCE}s",
                 warning,
@@ -2293,11 +2293,11 @@ async def test_a_fractional_usage_rounds_the_hold_and_the_preflight_up(self):
         assert kwargs["preflight"] == PreflightRequestBody(usage=3)
 
     async def test_a_reservation_ttl_above_the_cap_is_clamped(self):
-        client = _async_server_client(credit_leases=CreditLeaseConfig(default_reservation_ttl=7200.0))
+        client = _async_server_client(credit_leases=CreditLeaseConfig(default_reservation_ttl=90000.0))
         try:
             assert client._reservation_ttl == MAX_RESERVATION_TTL - RESERVATION_TTL_SKEW_ALLOWANCE
             warning = " ".join(str(call.args[0]) for call in client.logger.warning.call_args_list)
-            assert "one hour cap" in warning
+            assert "24 hour cap" in warning
             assert (
                 f"server-mode holds will be clamped to {MAX_RESERVATION_TTL - RESERVATION_TTL_SKEW_ALLOWANCE}s"
                 in warning
@@ -2995,16 +2995,16 @@ async def test_a_lease_gated_check_reports_one_flag_check_event(self):
     async def test_client_mode_keeps_a_reservation_ttl_past_the_server_cap(self):
         client = _async_lease_client(
             credit_leases=CreditLeaseConfig(
-                mode="client", default_lease_size=1000.0, default_reservation_ttl=7200.0
+                mode="client", default_lease_size=1000.0, default_reservation_ttl=90000.0
             )
         )
         try:
             # The TTL only drives the local sweeper here, so the server's cap
             # does not apply and nothing is clamped or warned about.
-            assert client._reservation_ttl == 7200.0
-            assert client._lease_manager.resolve_config("bilcr_inference").reservation_ttl == 7200.0
+            assert client._reservation_ttl == 90000.0
+            assert client._lease_manager.resolve_config("bilcr_inference").reservation_ttl == 90000.0
             warning = " ".join(str(call.args[0]) for call in client.logger.warning.call_args_list)
-            assert "one hour cap" not in warning
+            assert "24 hour cap" not in warning
         finally:
             await self._drain(client)