NBSOPSNEBIUS-25: report ExternalEndpointUnexpectedExit on vhost-serve…

…r crashes
ydb-platform · Jan 30, 2024 · 7057fee · 7057fee
1 parent 5097d0f
commit 7057fee
Show file tree

Hide file tree

Showing 14 changed files with 361 additions and 18 deletions.
diff --git a/cloud/blockstore/config/server.proto b/cloud/blockstore/config/server.proto
@@ -168,6 +168,10 @@ message TServerConfig
 
     // Path to vhost server executable.
     optional string VhostServerPath = 108;
+
+    // Additional vhost server command line arguments
+    // e.g. ["--pid-file", "/tmp/test-vhost-server-pid", "--verbose", "debug"]
+    repeated string VhostServerExtArgs = 109;
 }
 
 ////////////////////////////////////////////////////////////////////////////////

diff --git a/cloud/blockstore/libs/daemon/common/bootstrap.cpp b/cloud/blockstore/libs/daemon/common/bootstrap.cpp
@@ -407,6 +407,7 @@ void TBootstrapBase::Init()
                 ServerStats,
                 Executor,
                 Configs->ServerConfig->GetVhostServerPath(),
+                Configs->ServerConfig->GetVhostServerExtArgs(),
                 Configs->Options->SkipDeviceLocalityValidation
                     ? TString {}
                     : FQDNHostName(),

diff --git a/cloud/blockstore/libs/diagnostics/critical_events.h b/cloud/blockstore/libs/diagnostics/critical_events.h
@@ -51,6 +51,7 @@ namespace NCloud::NBlockStore {
     xxx(DiskRegistryDeviceNotFoundSoft)                                        \
     xxx(DiskRegistrySourceDiskNotFound)                                        \
     xxx(EndpointSwitchFailure)                                                 \
+    xxx(ExternalEndpointUnexpectedExit)                                        \
     xxx(DiskAgentSessionCacheUpdateError)                                      \
     xxx(DiskAgentSessionCacheRestoreError)                                     \
 // BLOCKSTORE_CRITICAL_EVENTS

diff --git a/cloud/blockstore/libs/endpoints_vhost/external_vhost_server.cpp b/cloud/blockstore/libs/endpoints_vhost/external_vhost_server.cpp
@@ -2,6 +2,7 @@
 #include "external_endpoint_stats.h"
 
 #include <cloud/blockstore/libs/common/device_path.h>
+#include <cloud/blockstore/libs/diagnostics/critical_events.h>
 #include <cloud/blockstore/libs/diagnostics/server_stats.h>
 #include <cloud/blockstore/libs/endpoints/endpoint_listener.h>
 
@@ -485,7 +486,9 @@ class TEndpoint final
                 break;
             }
 
-            // TODO: limiter
+            ReportExternalEndpointUnexpectedExit(TStringBuilder()
+                << "External endpoint " << Stats.DiskId << " " << Stats.ClientId
+                << " unexpectedly stopped: " << FormatError(error));
 
             auto process = RestartProcess();
             if (!process) {
@@ -943,6 +946,7 @@ IEndpointListenerPtr CreateExternalVhostEndpointListener(
     IServerStatsPtr serverStats,
     TExecutorPtr executor,
     TString binaryPath,
+    TVector<TString> extArgs,
     TString localAgentId,
     IEndpointListenerPtr fallbackListener)
 {
@@ -952,6 +956,8 @@ IEndpointListenerPtr CreateExternalVhostEndpointListener(
         TVector<TString> args,
         TVector<TString> cgroups)
     {
+        args.insert(args.end(), extArgs.begin(), extArgs.end());
+
         return std::make_shared<TEndpoint>(
             clientId,
             logging,

diff --git a/cloud/blockstore/libs/endpoints_vhost/external_vhost_server.h b/cloud/blockstore/libs/endpoints_vhost/external_vhost_server.h
@@ -37,6 +37,7 @@ IEndpointListenerPtr CreateExternalVhostEndpointListener(
     IServerStatsPtr serverStats,
     TExecutorPtr executor,
     TString binaryPath,
+    TVector<TString> extArgs,
     TString localAgentId,
     IEndpointListenerPtr fallbackListener);
 

diff --git a/cloud/blockstore/libs/server/config.cpp b/cloud/blockstore/libs/server/config.cpp
@@ -88,6 +88,7 @@ constexpr TDuration Seconds(int s)
         NCloud::NProto::ENDPOINT_STORAGE_KEYRING                              )\
     xxx(EndpointStorageDir,          TString,               {}                )\
     xxx(VhostServerPath,             TString,               {}                )\
+    xxx(VhostServerExtArgs,          TVector<TString>,      {}                )\
 // BLOCKSTORE_SERVER_CONFIG
 
 #define BLOCKSTORE_SERVER_DECLARE_CONFIG(name, type, value)                    \
@@ -116,19 +117,15 @@ template <>
 TVector<TString> ConvertValue(
     const google::protobuf::RepeatedPtrField<TString>& value)
 {
-    TVector<TString> v;
-    for (const auto& x : value) {
-        v.push_back(x);
-    }
-    return v;
+    return { value.begin(), value.end() };
 }
 
 template <>
 TVector<TCertificate> ConvertValue(
     const google::protobuf::RepeatedPtrField<NCloud::NProto::TCertificate>& value)
 {
     TVector<TCertificate> v;
-    for (const auto& x : value) {
+    for (const auto& x: value) {
         v.push_back({x.GetCertFile(), x.GetCertPrivateKeyFile()});
     }
     return v;

diff --git a/cloud/blockstore/libs/server/config.h b/cloud/blockstore/libs/server/config.h
@@ -124,6 +124,7 @@ class TServerAppConfig
     NCloud::NProto::EEndpointStorageType GetEndpointStorageType() const;
     TString GetEndpointStorageDir() const;
     TString GetVhostServerPath() const;
+    TVector<TString> GetVhostServerExtArgs() const;
 
     void Dump(IOutputStream& out) const override;
     void DumpHtml(IOutputStream& out) const override;

diff --git a/cloud/blockstore/public/sdk/python/client/base_client.py b/cloud/blockstore/public/sdk/python/client/base_client.py
@@ -41,6 +41,8 @@
     "execute_action",
     "kick_endpoint",
     "cms_action",
+    "update_disk_registry_config",
+    "describe_disk_registry_config",
 ]
 
 

diff --git a/cloud/blockstore/public/sdk/python/client/safe_client.py b/cloud/blockstore/public/sdk/python/client/safe_client.py
@@ -2,6 +2,8 @@
 
 import cloud.blockstore.public.sdk.python.protos as protos
 
+from google.protobuf.json_format import ParseDict
+
 from .error import _handle_errors
 
 
@@ -901,3 +903,69 @@ def cms_action(
             trace_id,
             request_timeout)
         return response
+
+    @_handle_errors
+    def update_disk_registry_config_async(
+            self,
+            config,
+            idempotence_id=None,
+            timestamp=None,
+            trace_id=None,
+            request_timeout=None):
+
+        request = ParseDict(config, protos.TUpdateDiskRegistryConfigRequest())
+
+        return self.__impl.update_disk_registry_config_async(
+            request,
+            idempotence_id,
+            timestamp,
+            trace_id,
+            request_timeout)
+
+    @_handle_errors
+    def update_disk_registry_config(
+            self,
+            config,
+            idempotence_id=None,
+            timestamp=None,
+            trace_id=None,
+            request_timeout=None):
+
+        request = ParseDict(config, protos.TUpdateDiskRegistryConfigRequest())
+
+        return self.__impl.update_disk_registry_config(
+            request,
+            idempotence_id,
+            timestamp,
+            trace_id,
+            request_timeout)
+
+    @_handle_errors
+    def describe_disk_registry_config_async(
+            self,
+            idempotence_id=None,
+            timestamp=None,
+            trace_id=None,
+            request_timeout=None):
+
+        return self.__impl.describe_disk_registry_config_async(
+            protos.TDescribeDiskRegistryConfigRequest(),
+            idempotence_id,
+            timestamp,
+            trace_id,
+            request_timeout)
+
+    @_handle_errors
+    def describe_disk_registry_config(
+            self,
+            idempotence_id=None,
+            timestamp=None,
+            trace_id=None,
+            request_timeout=None):
+
+        return self.__impl.describe_disk_registry_config(
+            protos.TDescribeDiskRegistryConfigRequest(),
+            idempotence_id,
+            timestamp,
+            trace_id,
+            request_timeout)