Merge upstream HEAD(bea04d9, 2024-05-21) [release] 0.2.130 (hail-is#14454)

Author: milo-hyben

.git-blame-ignore-revs

@@ -14,3 +14,5 @@ da2790242a40ec425a53a02707d261c893b264f7
 422edf6386616711ca70f87c455f76781ac925d4
 # replaces black formatting with ruff
 fa2ef0f2c76654d0c037ff6db60ccb8842fb8539
+# ruff lint python imports
+01a6a6a107faf204d4f5c20f8ae510d2c35518e9

Makefile

@@ -86,7 +86,6 @@ check-pip-requirements:
 		hail/python/dev \
 		gear \
 		web_common \
-		auth \
 		batch \
 		ci
 
@@ -98,7 +97,6 @@ check-linux-pip-requirements:
 		hail/python/dev \
 		gear \
 		web_common \
-		auth \
 		batch \
 		ci
 

batch/batch/batch.py

@@ -157,6 +157,8 @@ def job_record_to_dict(record: Dict[str, Any], name: Optional[str]) -> JobListEn
             'cost': coalesce(record.get('cost'), 0),
             'msec_mcpu': record['msec_mcpu'],
             'cost_breakdown': cost_breakdown,
+            'always_run': bool(record['always_run']),
+            'display_state': None,
         },
     )
 

batch/batch/cloud/gcp/worker/worker_api.py

@@ -1,6 +1,7 @@
 import base64
 import os
 import tempfile
+from contextlib import AsyncExitStack
 from typing import Dict, List
 
 import orjson
@@ -20,13 +21,13 @@
 class GCPWorkerAPI(CloudWorkerAPI):
     nameserver_ip = '169.254.169.254'
 
-    # async because GoogleSession must be created inside a running event loop
+    # async because ClientSession must be created inside a running event loop
     @staticmethod
     async def from_env() -> 'GCPWorkerAPI':
         project = os.environ['PROJECT']
         zone = os.environ['ZONE'].rsplit('/', 1)[1]
         worker_credentials = aiogoogle.GoogleInstanceMetadataCredentials()
-        http_session = httpx.ClientSession()
+        http_session = httpx.client_session()
         return GCPWorkerAPI(project, zone, worker_credentials, http_session)
 
     def __init__(
@@ -38,9 +39,15 @@ def __init__(
     ):
         self.project = project
         self.zone = zone
+
+        self._exit_stack = AsyncExitStack()
         self._http_session = http_session
-        self._metadata_server_client = aiogoogle.GoogleMetadataServerClient(http_session)
+        self._exit_stack.push_async_callback(self._http_session.close)
+
         self._compute_client = aiogoogle.GoogleComputeClient(project)
+        self._exit_stack.push_async_callback(self._compute_client.close)
+
+        self._metadata_server_client = aiogoogle.GoogleMetadataServerClient(http_session)
         self._gcsfuse_credential_files: Dict[str, str] = {}
         self._worker_credentials = worker_credentials
 
@@ -132,7 +139,7 @@ async def unmount_cloudfuse(self, mount_base_path_data: str):
             del self._gcsfuse_credential_files[mount_base_path_data]
 
     async def close(self):
-        await self._compute_client.close()
+        await self._exit_stack.aclose()
 
     def __str__(self):
         return f'project={self.project} zone={self.zone}'

batch/batch/front_end/front_end.py

@@ -2449,6 +2449,11 @@ async def ui_batch(request, userdata, batch_id):
     for j in jobs:
         j['duration'] = humanize_timedelta_msecs(j['duration'])
         j['cost'] = cost_str(j['cost'])
+        j['display_state'] = (
+            f"{j['state']} (always run)"
+            if j['always_run'] and j['state'] not in {'Success', 'Failed', 'Error'}
+            else j['state']
+        )
     batch['jobs'] = jobs
 
     batch['cost'] = cost_str(batch['cost'])

batch/batch/front_end/templates/batch.html

@@ -91,7 +91,7 @@ <h2>Jobs</h2>
             {{ job['name'] }}
             {% endif %}
           </td>
-          <td>{{ job['state'] }}</td>
+          <td>{{ job['display_state'] }}</td>
           <td>
             {% if 'exit_code' in job and job['exit_code'] is not none %}
             {{ job['exit_code'] }}

batch/batch/front_end/templates/billing.html

@@ -34,19 +34,19 @@ <h1>Billing</h1>
         </div>
     </div>
 
-    <h2>Total Cost</h2>
+    <h2>Total Spend</h2>
     <ul>
         <li>{{ total_cost }}</li>
     </ul>
 
     {% if is_developer %}
-    <h2>Cost by Billing Project</h2>
+    <h2>Spend by Billing Project</h2>
     <div class='flex-col' style="overflow: auto;">
         <table class="data-table" id="billing_by_project">
             <thead>
             <tr>
                 <th>Billing Project</th>
-                <th>Cost</th>
+                <th>Spend</th>
             </tr>
             </thead>
             <tbody>
@@ -60,13 +60,13 @@ <h2>Cost by Billing Project</h2>
         </table>
     </div>
 
-    <h2>Cost by User</h2>
+    <h2>Spend by User</h2>
     <div class='flex-col' style="overflow: auto;">
         <table class="data-table" id="billing_by_user">
             <thead>
             <tr>
                 <th>User</th>
-                <th>Cost</th>
+                <th>Spend</th>
             </tr>
             </thead>
             <tbody>
@@ -81,14 +81,14 @@ <h2>Cost by User</h2>
     </div>
     {% endif %}
 
-    <h2>Cost by Billing Project and User</h2>
+    <h2>Spend by Billing Project and User</h2>
     <div class='flex-col' style="overflow: auto;">
         <table class="data-table" id="billing_by_project_user">
             <thead>
             <tr>
                 <th>Billing Project</th>
                 <th>User</th>
-                <th>Cost</th>
+                <th>Spend</th>
             </tr>
             </thead>
             <tbody>

batch/batch/front_end/templates/job.html

@@ -13,6 +13,7 @@ <h2>Properties</h2>
   <li>Exit Code: {% if 'exit_code' in job and job['exit_code'] is not none %}{{ job['exit_code'] }}{% endif %}</li>
   <li>Duration: {% if 'duration' in job and job['duration'] is not none %}{{ job['duration'] }}{% endif %}</li>
   <li>Cost: {% if 'cost' in job and job['cost'] is not none %}{{ job['cost'] }}{% endif %}</li>
+  <li>Always Run: {% if 'always_run' in job and job['always_run'] is not none %}{{ job['always_run'] }}{% endif %}</li>
 </ul>
 
 <h2>Attributes</h2>

batch/batch/utils.py

@@ -157,8 +157,7 @@ async def query_billing_projects_with_cost(db, user=None, billing_project=None)
   ) AS usage_t
   LEFT JOIN resources ON resources.resource_id = usage_t.resource_id
 ) AS cost_t ON TRUE
-{where_condition}
-LOCK IN SHARE MODE;
+{where_condition};
 """
 
     billing_projects = []

batch/batch/worker/worker.py

@@ -2936,7 +2936,7 @@ def __repr__(self):
 
 
 class Worker:
-    def __init__(self, client_session: httpx.ClientSession):
+    def __init__(self):
         self.active = False
         self.cores_mcpu = CORES * 1000
         self.last_updated = time_msecs()
@@ -2948,7 +2948,7 @@ def __init__(self, client_session: httpx.ClientSession):
         self.task_manager = aiotools.BackgroundTaskManager()
         os.makedirs('/hail-jars/', exist_ok=True)
         self.jar_download_locks: Dict[str, asyncio.Lock] = defaultdict(asyncio.Lock)
-        self.client_session = client_session
+        self.client_session = httpx.client_session()
 
         self.image_data: Dict[str, ImageData] = defaultdict(ImageData)
         self.image_data[BATCH_WORKER_IMAGE_ID] += 1
@@ -3452,7 +3452,7 @@ async def async_main():
     network_allocator = NetworkAllocator(network_allocator_task_manager)
     await network_allocator.reserve()
 
-    worker = Worker(httpx.client_session())
+    worker = Worker()
     try:
         async with AsyncExitStack() as cleanup:
             cleanup.push_async_callback(docker.close)

batch/test/test_batch.py

@@ -1,3 +1,4 @@
+import asyncio
 import collections
 import os
 import secrets
@@ -1497,18 +1498,23 @@ def test_pool_standard_instance_cheapest(client: BatchClient):
 
 # Transitively is not valid for terra
 @skip_in_azure
-@pytest.mark.timeout(10 * 60)
-def test_gpu_accesibility_g2(client: BatchClient):
-    b = create_batch(client)
+async def test_gpu_accesibility_g2(client: BatchClient):
+    b = create_batch(client)._async_batch
     resources = {'machine_type': "g2-standard-4", 'storage': '100Gi'}
     j = b.create_job(
         os.environ['HAIL_GPU_IMAGE'],
         ['python3', '-c', 'import torch; assert torch.cuda.is_available()'],
         resources=resources,
     )
-    b.submit()
-    status = j.wait()
-    assert status['state'] == 'Success', str((status, b.debug_info()))
+    await b.submit()
+    try:
+        status = await asyncio.wait_for(j.wait(), timeout=5 * 60)
+        assert status['state'] == 'Success', str((status, b.debug_info()))
+    except asyncio.TimeoutError:
+        # G2 instances are not always available within a time window
+        # acceptable for CI. This test is permitted to time out
+        # but not otherwise fail
+        pass
 
 
 def test_job_private_instance_preemptible(client: BatchClient):

benchmark/Makefile

@@ -1,8 +1,6 @@
 include ../config.mk
 include ../hail/version.mk
 
-SHORT_REVISION := $(shell git rev-parse --short=12 HEAD)
-
 HAIL_PYTHON3 ?= python3
 PIP := $(HAIL_PYTHON3) -m pip
 
@@ -53,21 +51,21 @@ pushed_image: image
 
 BENCHMARK_ITERS ?= 3
 BENCHMARK_REPLICATES ?= 5
-HAIL_WHEEL_DESCRIPTOR ?= $(HAIL_PIP_VERSION)-$(SHORT_REVISION)
 BENCHMARK_BUCKET ?= gs://hail-benchmarks-2
 .PHONY: submit
 submit: pushed_image install
 	@echo Using pushed image `cat pushed_image`
 	$(HAIL_PYTHON3) scripts/benchmark_in_batch.py \
 		`cat pushed_image` \
 		$(BENCHMARK_BUCKET)/$(shell whoami) \
-		$(HAIL_WHEEL_DESCRIPTOR) \
+		$(HAIL_VERSION) \
 		$(BENCHMARK_REPLICATES) \
-		$(BENCHMARK_ITERS)
+		$(BENCHMARK_ITERS) \
+		run
 
 clean: cleanup_image
 	rm -rf python/dist/*
 	rm -rf python/build/*
-	rm -r $(HAIL_BENCHMARK_VERSION_FILE)
+	rm -f $(HAIL_BENCHMARK_VERSION_FILE)
 
 FORCE: