DockerLogin retries

src/Agent.Sdk/Knob/AgentKnobs.cs

@@ -61,6 +61,15 @@ public class AgentKnobs
             new EnvironmentKnobSource("AZP_AGENT_USE_HOST_GROUP_ID"),
             new BuiltInDefaultKnobSource("true"));
 
+        public const string DockerActionRetriesVariableName = "VSTSAGENT_DOCKER_ACTION_RETRIES";
+
+        public static readonly Knob DockerActionRetries = new Knob(
+            nameof(DockerActionRetries),
+            "When enabled, the agent retries docker steps if failed",
+            new RuntimeKnobSource(DockerActionRetriesVariableName),
+            new EnvironmentKnobSource(DockerActionRetriesVariableName),
+            new BuiltInDefaultKnobSource("false"));
+
         // Directory structure
         public static readonly Knob AgentToolsDirectory = new Knob(
             nameof(AgentToolsDirectory),

src/Agent.Worker/Container/DockerCommandManager.cs

@@ -57,7 +57,9 @@ public override void Initialize(IHostContext hostContext)
         public async Task<DockerVersion> DockerVersion(IExecutionContext context)
         {
             ArgUtil.NotNull(context, nameof(context));
-            string serverVersionStr = (await ExecuteDockerCommandAsync(context, "version", "--format '{{.Server.APIVersion}}'")).FirstOrDefault();
+            var action = new Func<Task<List<string>>>(async () => await ExecuteDockerCommandAsync(context, "version", "--format '{{.Server.APIVersion}}'"));
+            const string command = "Docker version";
+            string serverVersionStr = (await ExecuteDockerCommandAsyncWithRetries(context, action, command)).FirstOrDefault();
             ArgUtil.NotNullOrEmpty(serverVersionStr, "Docker.Server.Version");
             context.Output($"Docker daemon API version: {serverVersionStr}");
 
@@ -98,12 +100,14 @@ public async Task<int> DockerLogin(IExecutionContext context, string server, str
             ArgUtil.NotNull(username, nameof(username));
             ArgUtil.NotNull(password, nameof(password));
 
-            if (PlatformUtil.RunningOnWindows)
-            {
+            var action = new Func<Task<int>>(async () => PlatformUtil.RunningOnWindows
                 // Wait for 17.07 to switch using stdin for docker registry password.
-                return await ExecuteDockerCommandAsync(context, "login", $"--username \"{username}\" --password \"{password.Replace("\"", "\\\"")}\" {server}", new List<string>() { password }, context.CancellationToken);
-            }
-            return await ExecuteDockerCommandAsync(context, "login", $"--username \"{username}\" --password-stdin {server}", new List<string>() { password }, context.CancellationToken);
+                ? await ExecuteDockerCommandAsync(context, "login", $"--username \"{username}\" --password \"{password.Replace("\"", "\\\"")}\" {server}", new List<string>() { password }, context.CancellationToken)
+                : await ExecuteDockerCommandAsync(context, "login", $"--username \"{username}\" --password-stdin {server}", new List<string>() { password }, context.CancellationToken)
+            );
+
+            const string command = "Docker login";
+            return await ExecuteDockerCommandAsyncWithRetries(context, action, command);
         }
 
         public async Task<int> DockerLogout(IExecutionContext context, string server)
@@ -119,7 +123,9 @@ public async Task<int> DockerPull(IExecutionContext context, string image)
             ArgUtil.NotNull(context, nameof(context));
             ArgUtil.NotNull(image, nameof(image));
 
-            return await ExecuteDockerCommandAsync(context, "pull", image, context.CancellationToken);
+            var action = new Func<Task<int>>(async () => await ExecuteDockerCommandAsync(context, "pull", image, context.CancellationToken));
+            const string command = "Docker pull";
+            return await ExecuteDockerCommandAsyncWithRetries(context, action, command);
         }
 
         public async Task<string> DockerCreate(IExecutionContext context, ContainerInfo container)
@@ -194,7 +200,9 @@ public async Task<int> DockerStart(IExecutionContext context, string containerId
             ArgUtil.NotNull(context, nameof(context));
             ArgUtil.NotNull(containerId, nameof(containerId));
 
-            return await ExecuteDockerCommandAsync(context, "start", containerId, context.CancellationToken);
+            var action = new Func<Task<int>>(async () => await ExecuteDockerCommandAsync(context, "start", containerId, context.CancellationToken));
+            const string command = "Docker start";
+            return await ExecuteDockerCommandAsyncWithRetries(context, action, command);
         }
 
         public async Task<int> DockerRemove(IExecutionContext context, string containerId)
@@ -443,5 +451,70 @@ await processInvoker.ExecuteAsync(
 
             return output;
         }
+
+        private static async Task<int> ExecuteDockerCommandAsyncWithRetries(IExecutionContext context, Func<Task<int>> action, string command)
+        {
+            bool dockerActionRetries = AgentKnobs.DockerActionRetries.GetValue(context).AsBoolean();
+            context.Output($"DockerActionRetries variable value: {dockerActionRetries}");
+
+            int retryCount = 0;
+            int exitCode = 0;
+            const int maxRetries = 3;
+            TimeSpan delayInSeconds = TimeSpan.FromSeconds(10);
+
+            while (retryCount < maxRetries)
+            {
+                exitCode = await action();
+
+                if (exitCode == 0 || !dockerActionRetries)
+                {
+                    break;
+                }
+
+                context.Warning($"{command} failed with exit code {exitCode}, back off {delayInSeconds} seconds before retry.");
+                await Task.Delay(delayInSeconds);
+                retryCount++;
+            }
+
+            return exitCode;
+        }
+
+        private static async Task<List<string>> ExecuteDockerCommandAsyncWithRetries(IExecutionContext context, Func<Task<List<string>>> action, string command)
+        {
+            bool dockerActionRetries = AgentKnobs.DockerActionRetries.GetValue(context).AsBoolean();
+            context.Output($"DockerActionRetries variable value: {dockerActionRetries}");
+
+            int retryCount = 0;
+            List<string> output = new List<string>();
+            const int maxRetries = 3;
+            TimeSpan delayInSeconds = TimeSpan.FromSeconds(10);
+
+            while (retryCount <= maxRetries)
+            {
+                try
+                {
+                    output = await action();
+                }
+                catch (ProcessExitCodeException)
+                {
+                    if (!dockerActionRetries || retryCount == maxRetries)
+                    {
+                        throw;
+                    }
+
+                    context.Warning($"{command} failed, back off {delayInSeconds} seconds before retry.");
+                    await Task.Delay(delayInSeconds);
+                }
+
+                retryCount++;
+
+                if (output != null && output.Count != 0)
+                {
+                    break;
+                }
+            }
+
+            return output;
+        }
     }
 }

src/Agent.Worker/ContainerOperationProvider.cs

@@ -317,7 +317,12 @@ private async Task PullContainerAsync(IExecutionContext executionContext, Contai
                 ArgUtil.NotNullOrEmpty(username, nameof(username));
                 ArgUtil.NotNullOrEmpty(password, nameof(password));
 
-                int loginExitCode = await _dockerManger.DockerLogin(executionContext, registryServer, username, password);
+                int loginExitCode = await _dockerManger.DockerLogin(
+                    executionContext,
+                    registryServer,
+                    username,
+                    password);
+
                 if (loginExitCode != 0)
                 {
                     throw new InvalidOperationException($"Docker login fail with exit code {loginExitCode}");
@@ -338,29 +343,11 @@ private async Task PullContainerAsync(IExecutionContext executionContext, Contai
                         }
                     }
 
-                    // Pull down docker image with retry up to 3 times
-                    int retryCount = 0;
-                    int pullExitCode = 0;
-                    while (retryCount < 3)
-                    {
-                        pullExitCode = await _dockerManger.DockerPull(executionContext, container.ContainerImage);
-                        if (pullExitCode == 0)
-                        {
-                            break;
-                        }
-                        else
-                        {
-                            retryCount++;
-                            if (retryCount < 3)
-                            {
-                                var backOff = BackoffTimerHelper.GetRandomBackoff(TimeSpan.FromSeconds(1), TimeSpan.FromSeconds(10));
-                                executionContext.Warning($"Docker pull failed with exit code {pullExitCode}, back off {backOff.TotalSeconds} seconds before retry.");
-                                await Task.Delay(backOff);
-                            }
-                        }
-                    }
+                    int pullExitCode = await _dockerManger.DockerPull(
+                        executionContext,
+                        container.ContainerImage);
 
-                    if (retryCount == 3 && pullExitCode != 0)
+                    if (pullExitCode != 0)
                     {
                         throw new InvalidOperationException($"Docker pull failed with exit code {pullExitCode}");
                     }