diff --git a/design/velero-debug.md b/design/velero-debug.md index 6b2454c2266..523c56cc749 100644 --- a/design/velero-debug.md +++ b/design/velero-debug.md @@ -21,12 +21,15 @@ A dedicated subcommand can lower this effort and reduce back-and-forth between u With the introduction of the new command `velero debug`, the command would download all of the following information: - velero deployment logs - restic DaemonSet logs -- Plugin logs - need clarification for vSphere plugin see open quetions -- Resource and log of the backup and restore, if specified in the param -- Resources: - - BackupStorageLocation - - PodVolumeBackups - - PodVolumeRestores +- plugin logs +- All the resources in the group `velero.io` that are created such as: + - Backup + - Restore + - BackupStorageLocation + - PodVolumeBackup + - PodVolumeRestore + - *etc ...* +- Log of the backup and restore, if specified in the param A project called `crash-diagnostics` (or `crashd`) (https://github.com/vmware-tanzu/crash-diagnostics) implements the Kubernetes API queries and provides Starlark scripting language to abstract details, and collect the information into a local copy. It can be used as a standalone CLI executing a Starlark script file. With the capabilities of embedding files in Go 1.16, we can define a Starlark script gathering the necessary information, embed the script at build time, then the velero debug command will invoke `crashd`, passing in the script’s text contents. @@ -36,29 +39,31 @@ With the capabilities of embedding files in Go 1.16, we can define a Starlark sc The Starlark script to be called by crashd: ```python -def capture_backup_logs(): +def capture_backup_logs(namespace): if args.backup: - kube_capture(what="objects", kinds=['backups'], names=[args.backup]) - backupLogsCmd = "velero backup logs {}".format(args.backup) - capture_local(cmd=backupLogsCmd) -def capture_restore_logs(): + log("Collecting log for backup: {}".format(args.backup)) + backupLogsCmd = "velero --namespace={} backup logs {}".format(namespace, args.backup) + capture_local(cmd=backupLogsCmd, file_name="backup_{}.log".format(args.backup)) +def capture_restore_logs(namespace): if args.restore: - kube_capture(what="objects", kinds=['restores'], names=[args.restore]) - restoreLogsCmd = "velero restore logs {}".format(args.restore) - capture_local(cmd=restoreLogsCmd) + log("Collecting log for restore: {}".format(args.restore)) + restoreLogsCmd = "velero --namespace={} restore logs {}".format(namespace, args.restore) + capture_local(cmd=restoreLogsCmd, file_name="restore_{}.log".format(args.restore)) ns = args.namespace if args.namespace else "velero" -basedir = args.basedir if args.basedir else os.home output = args.output if args.output else "bundle.tar.gz" # Working dir for writing during script execution -crshd = crashd_config(workdir="{0}/velero-bundle".format(basedir)) -set_defaults(kube_config(path=args.kubeconfig)) -capture_local(cmd="velero version -n {}".format(ns)) -capture_backup_logs() -capture_restore_logs() +crshd = crashd_config(workdir="./velero-bundle") +set_defaults(kube_config(path=args.kubeconfig, cluster_context=args.kubecontext)) +log("Collecting velero resources in namespace: {}". format(ns)) +kube_capture(what="objects", namespaces=[ns], groups=['velero.io']) +capture_local(cmd="velero version -n {}".format(ns), file_name="version.txt") +log("Collecting velero deployment logs in namespace: {}". format(ns)) kube_capture(what="logs", namespaces=[ns]) -kube_capture(what="objects", namespaces=[ns], kinds=['backupstoragelocations', 'podvolumebackups', 'podvolumerestores']) +capture_backup_logs(ns) +capture_restore_logs(ns) archive(output_file=output, source_paths=[crshd.workdir]) +log("Generated debug information bundle: {}".format(output)) ``` The sample command to trigger the script via crashd: ```shell @@ -68,41 +73,32 @@ The sample command to trigger the script via crashd: To trigger the script in `velero debug`, in the package `pkg/cmd/cli/debug` a struct `option` will be introduced ```go type option struct { - // workdir for crashd will be $baseDir/tmp/crashd - baseDir string - // the namespace where velero server is installed - namespace string - // the absolute path for the log bundle to be generated - outputPath string - // the absolute path for the kubeconfig file that will be read by crashd for calling K8S API - kubeconfigPath string - // optional, the name of the backup resource whose log will be packaged into the debug bundle - backup string - // optional, the name of the restore resource whose log will be packaged into the debug bundle - restore string + // workdir for crashd will be $baseDir/velero-debug + baseDir string + // the namespace where velero server is installed + namespace string + // the absolute path for the log bundle to be generated + outputPath string + // the absolute path for the kubeconfig file that will be read by crashd for calling K8S API + kubeconfigPath string + // the kubecontext to be used for calling K8S API + kubeContext string + // optional, the name of the backup resource whose log will be packaged into the debug bundle + backup string + // optional, the name of the restore resource whose log will be packaged into the debug bundle + restore string + // optional, it controls whether to print the debug log messages when calling crashd + verbose bool } ``` -The code will consolidate the input parameters and execution context of the `velero` CLI to form the option struct, which can be transformed into the `args` string for `crashd` -### kubeconfig -When it comes to accessing the API of k8s, `crashd` has a limitation that it can only accept a path of kubeconfig file, without customizing the `context`, and it does not honor the environment variables such as `KUBECONFIG`. `velero` does honor the environment variables and allow user to customize the path to kubeconfig and the `context` -There are two ways to make crashd have consistent behavior as velero in terms of getting the kube configuration: -1. Modify crashd to make it honor the environment variable and allow user to set context while calling k8s APIs. This is a preferred approach and it does make `crashd` better, but it may take longer time because we need to convince the maintainers of `crashd`, and double check the change will not break their current use cases. - There are 2 issues opened: - https://github.com/vmware-tanzu/crash-diagnostics/issues/208 - https://github.com/vmware-tanzu/crash-diagnostics/issues/122 - I'll try to contact the maintainers of `crashd` to see the feasibility for velero v1.7 -2. Before calling the `crashd` script velero CLI will use `client-go` to generate a temp `kubeconfig` file honoring the environment variable and global flags, and pass it to crashd. Although there’s no permission elevation and the temp file will be removed, there’s still some security concern because the temp file is accessible by other programs before it’s deleted, or it may not be deleted if an error happens. - -Therefore, we should consider `option 1` the better choice, and see `option 2` as the backup. +The code will consolidate the input parameters and execution context of the `velero` CLI to form the option struct, which can be transformed into the `argsMap` that can be used when calling the func `exec.Execute` in `crashd`: +https://github.com/vmware-tanzu/crash-diagnostics/blob/v0.3.4/exec/executor.go#L17 ## Alternatives Considered The collection could be done via the kubernetes client-go API, but such integration is not necessarily trivial to implement, therefore, `crashd` is preferred approach - ## Security Considerations -- The current released version of `crashd` depends on `client-go v0.19.0` which has a known CVE, we need to make sure that when it’s compiled into velero it uses the version that has the CVE fixed. We should write a PR or push crashd maintainer to fix the CVE-2021-3121 in 0.19.0 -- The starlark script will be embedded into the velero binary, so there’s little risk that the script will be modified before being called. -- There may be minor security issues if we choose to create a temp `kubeconfig` file for `crashd` and remove it afterwards. If we have to choose this option, we need to review it with security experts to better understand the risks. +- The starlark script will be embedded into the velero binary, and the byte slice will be passed to the `exec.Execute` func directly, so there’s little risk that the script will be modified before being executed. ## Compatibility As the `crashd` project evolves the behavior of the internal functions used in the Starlark script may change. We’ll ensure the correctness of the script via regular E2E tests. @@ -115,6 +111,5 @@ As the `crashd` project evolves the behavior of the internal functions used in t 4. Add E2E test case ## Open Questions -- **Log collection for vsphere plugin:** Per the design of vsphere plugin: https://github.com/vmware-tanzu/velero-plugin-for-vsphere#architecture when user backup resource on a guest cluster the code in component in the supervisor cluster may be called. Per discussion in v1.7 we will only support collecting logs of process running in one k8s cluster. In terms of implementation, we will do investigate the possibility to call extra script in crashd and ask vsphere plugin developer to provide a script to do the log collection, but the details remain TBD. - **Command dependencies:** In the Starlark script, for collecting version info and backup logs, it calls the `velero backup logs` and `velero version`, which makes the call stack like velero debug -> crashd -> velero xxx. We need to make sure this works under different PATH settings. -- **Progress and error handling:** The log collection may take a relatively long time, log messages should be printed to indicate the progress when different items are being downloaded and packaged. Additionally, when an error happens, we need to double check if it’s omitted by crashd. +- **Progress and error handling:** The log collection may take a relatively long time, log messages should be printed to indicate the progress when different items are being downloaded and packaged. Additionally, when an error happens, `crashd` may omit some errors, so before the script is executed we'll do some validation and make sure the `debug` command fail early if some parameters are incorrect.