google-nomulus/docs/app-engine-architecture.md

App Engine architecture

This document contains information on the overall architecture of the Domain Registry project as it is implemented in App Engine.

Services

The Domain Registry contains three services, which were previously called modules in earlier versions of App Engine. The services are: default (also called front-end), backend, and tools. Each service runs independently in a lot of ways, including that they can be upgraded individually, their log outputs are separate, and their servers and configured scaling are separate as well.

Default service

The default service is responsible for all registrar-facing EPP command traffic, all user-facing WHOIS and RDAP traffic, and the admin and registrar web consoles, and is thus the most important service. If the service has any problems and goes down or stops servicing requests in a timely manner, it will begin to impact users immediately. Requests to the default service are handled by the FrontendServlet, which provides all of the endpoints exposed in FrontendRequestComponent.

Backend service

The backend service is responsible for executing all regularly scheduled background tasks (using cron) as well as all asynchronous tasks. Requests to the backend service are handled by the BackendServlet, which provides all of the endpoints exposed in BackendRequestComponent. These include tasks for generating/exporting RDE, syncing the trademark list from TMDB, exporting backups, writing out DNS updates, handling asynchronous contact and host deletions, writing out commit logs, exporting metrics to BigQuery, and many more. Issues in the backend service will not immediately be apparent to end users, but the longer it is down, the more obvious it will become that user-visible tasks such as DNS and deletion are not being handled in a timely manner.

The backend service is also where all MapReduces run, which includes some of the aforementioned tasks such as RDE and asynchronous resource deletion, as well as any one-off data migration MapReduces. Consequently, the backend service should be sized to support not just the normal ongoing DNS load but also the load incurred by MapReduces, both scheduled (such as RDE) and on-demand (asynchronous contact/host deletion).

Tools service

The tools service is responsible for servicing requests from the registry_tool command line tool, which provides administrative-level functionality for developers and tech support employees of the registry. It is thus the least critical of the three services. Requests to the tools service are handled by the ToolsServlet, which provides all of the endpoints exposed in ToolsRequestComponent. Some example functionality that this service provides includes the server-side code to update premium lists, run EPP commands from the tool, and manually modify contacts/hosts/domains/and other resources. Problems with the tools service are not visible to users.

Task queues

Task queues in App Engine provide an asynchronous way to enqueue tasks and then execute them on some kind of schedule. There are two types of queues, push queues and pull queues. Tasks in push queues are always executing up to some throttlable limit. Tasks in pull queues remain there indefinitely until the queue is polled by code that is running for some other reason. Essentially, push queues run their own tasks while pull queues just enqueue data that is used by something else. Many other parts of App Engine are implemented using task queues. For example, App Engine cron adds tasks to push queues at regularly scheduled intervals, and the MapReduce framework adds tasks for each phase of the MapReduce algorithm.

The Domain Registry project uses a particular pattern of paired push/pull queues that is worth explaining in detail. Push queues are essential because App Engine's architecture does not support long-running background processes, and so push queues are thus the fundamental building block that allows asynchronous and background execution of code that is not in response to incoming web requests. However, they also have limitations in that they do not allow batch processing or grouping. That's where the pull queue comes in. Regularly scheduled tasks in the push queue will, upon execution, poll the corresponding pull queue for a specified number of tasks and execute them in a batch. This allows the code to execute in the background while taking advantage of batch processing.

Particulars on the task queues in use by the Domain Registry project are specified in the queue.xml file. Note that many push queues have a direct one-to-one correspondence with entries in cron.xml because they need to be fanned-out on a per-TLD or other basis (see the Cron section below for more explanation). The exact queue that a given cron task will use is passed as the query string parameter "queue" in the url specification for the cron task.

Here are the task queues in use by the system. All are push queues unless explicitly marked as otherwise.

• bigquery-streaming-metrics -- Queue for metrics that are asynchronously streamed to BigQuery in the Metrics class. Tasks are enqueued during EPP flows in EppController. This means that there is a lag of a few seconds to a few minutes between when metrics are generated and when they are queryable in BigQuery, but this is preferable to slowing all EPP flows down and blocking them on BigQuery streaming.
• brda -- Queue for tasks to upload weekly Bulk Registration Data Access (BRDA) files to a location where they are available to ICANN. The RdeStagingReducer (part of the RDE MapReduce) creates these tasks at the end of generating an RDE dump.
• delete-commits -- Cron queue for tasks to regularly delete commit logs that are more than thirty days stale. These tasks execute the DeleteOldCommitLogsAction.
• dns-cron (cron queue) and dns-pull (pull queue) -- A push/pull pair of queues. Cron regularly enqueues tasks in dns-cron each minute, which are then executed by ReadDnsQueueAction, which leases a batch of tasks from the pull queue, groups them by TLD, and writes them as a single task to dns-publish to be published to the configured DNS writer for the TLD.
• dns-publish -- Queue for batches of DNS updates to be pushed to DNS writers.
• export-bigquery-poll -- Queue for tasks to query the success/failure of a given BigQuery export job. Tasks are enqueued by BigqueryPollJobAction.
• export-commits -- Queue for tasks to export commit log checkpoints. Tasks are enqueued by CommitLogCheckpointAction (which is run every minute by cron) and executed by ExportCommitLogDiffAction.
• export-reserved-terms -- Cron queue for tasks to export the list of reserved terms for each TLD. The tasks are executed by ExportReservedTermsAction.
• export-snapshot -- Cron and push queue for tasks to load a Datastore snapshot that was stored in Google Cloud Storage and export it to BigQuery. Tasks are enqueued by both cron and CheckSnapshotServlet and are executed by both ExportSnapshotServlet and LoadSnapshotAction.
• export-snapshot-poll -- Queue for tasks to check that a Datastore snapshot has been successfully uploaded to Google Cloud Storage (this is an asynchronous background operation that can take an indeterminate amount of time). Once the snapshot is successfully uploaded, it is imported into BigQuery. Tasks are enqueued by ExportSnapshotServlet and executed by CheckSnapshotServlet.
• export-snapshot-update-view -- Queue for tasks to update the BigQuery views to point to the most recently uploaded snapshot. Tasks are enqueued by LoadSnapshotAction and executed by UpdateSnapshotViewAction.
• flows-async -- Queue for asynchronous tasks that are enqueued during EPP command flows. Currently all of these tasks correspond to invocations of any of the following three MapReduces: DnsRefreshForHostRenameAction, DeleteHostResourceAction, or DeleteContactResourceAction.
• group-members-sync -- Cron queue for tasks to sync registrar contacts (not domain contacts!) to Google Groups. Tasks are executed by SyncGroupMembersAction.
• load[0-9] -- Queues used to load-test the system by LoadTestAction. These queues don't need to exist except when actively running load tests (which is not recommended on production environments). There are ten of these queues to provide simple sharding, because the Domain Registry system is capable of handling significantly more Queries Per Second than the highest throttle limit available on task queues (which is 500 qps).
• lordn-claims and lordn-sunrise -- Pull queues for handling LORDN exports. Tasks are enqueued synchronously during EPP commands depending on whether the domain name in question has a claims notice ID.
• marksdb -- Queue for tasks to verify that an upload to NORDN was successfully received and verified. These tasks are enqueued by NordnUploadAction following an upload and are executed by NordnVerifyAction.
• nordn -- Cron queue used for NORDN exporting. Tasks are executed by NordnUploadAction, which pulls LORDN data from the lordn-claims and lordn-sunrise pull queues (above).
• rde-report -- Queue for tasks to upload RDE reports to ICANN following successful upload of full RDE files to the escrow provider. Tasks are enqueued by RdeUploadAction and executed by RdeReportAction.
• rde-upload -- Cron queue for tasks to upload already-generated RDE files from Cloud Storage to the escrow provider. Tasks are executed by RdeUploadAction.
• sheet -- Queue for tasks to sync registrar updates to a Google Sheets spreadsheet. Tasks are enqueued by RegistrarServlet when changes are made to registrar fields and are executed by SyncRegistrarsSheetAction.

Environments

The domain registry codebase comes pre-configured with support for a number of different environments, all of which are used in Google's registry system. Other registry operators may choose to user more or fewer environments, depending on their needs.

The different environments are specified in RegistryEnvironment. Most correspond to a separate App Engine app except for UNITTEST and LOCAL, which by their nature do not use real environments running in the cloud. The recommended naming scheme for the App Engine apps that has the best possible compatibility with the codebase and thus requires the least configuration is to pick a name for the production app and then suffix it for the other environments. E.g., if the production app is to be named 'registry-platform', then the sandbox app would be named 'registry-platform-sandbox'.

The full list of environments supported out-of-the-box, in descending order from real to not, is:

• PRODUCTION -- The real production environment that is actually running live TLDs. Since the Domain Registry is a shared registry platform, there need only ever be one of these.
• SANDBOX -- A playground environment for external users to test commands in without the possibility of affecting production data. This is the environment new registrars go through OT&E in. Sandbox is also useful as a final sanity check to push a new prospective build to and allow it to "bake" before pushing it to production.
• QA -- An internal environment used by business users to play with and sign off on new features to be released. This environment can be pushed to frequently and is where manual testers should be spending the majority of their time.
• CRASH -- Another environment similar to QA, except with no expectations of data preservation. Crash is used for testing of backup/restore (which brings the entire system down until it is completed) without affecting the QA environment.
• ALPHA -- The developers' playground. Experimental builds are routinely pushed here in order to test them on a real app running on App Engine. You may end up wanting multiple environments like Alpha if you regularly experience contention (i.e. developers being blocked from testing their code on Alpha because others are already using it).
• LOCAL -- A fake environment that is used when running the app locally on a simulated App Engine instance.
• UNITTEST -- A fake environment that is used in unit tests, where everything in the App Engine stack is simulated or mocked.

Release process

The following is a recommended release process based on Google's several years of experience running a production registry using this codebase.

1. Developers write code and associated unit tests verifying that the new code works properly.
2. New features or potentially risky bug fixes are pushed to Alpha and tested by the developers before being committed to the source code repository.
3. New builds are cut and first pushed to Sandbox.
4. Once a build has been running successfully in Sandbox for a day with no errors, it can be pushed to Production.
5. Repeat once weekly, or potentially more often.

Cron tasks

All cron tasks are specified in cron.xml files, with one per environment. There are more tasks that execute in Production than in other environments, because tasks like uploading RDE dumps are only done for the live system. Cron tasks execute on the backend service.

Most cron tasks use the TldFanoutAction which is accessed via the /_dr/cron/fanout URL path. This action, which is run by the BackendServlet on the backend service, fans out a given cron task for each TLD that exists in the registry system, using the queue that is specified in the cron.xml entry. Because some tasks may be computationally intensive and could risk spiking system latency if all start executing immediately at the same time, there is a jitterSeconds parameter that spreads out tasks over the given number of seconds. This is used with DNS updates and commit log deletion.

The reason the TldFanoutAction exists is that a lot of tasks need to be done separately for each TLD, such as RDE exports and NORDN uploads. It's simpler to have a single cron entry that will create tasks for all TLDs than to have to specify a separate cron task for each action for each TLD (though that is still an option). Task queues also provide retry semantics in the event of transient failures that a raw cron task does not. This is why there are some tasks that do not fan out across TLDs that still use TldFanoutAction -- it's so that the tasks retry in the face of transient errors.

The full list of URL parameters to TldFanoutAction that can be specified in cron.xml is:

• endpoint -- The path of the action that should be executed (see web.xml).
• queue -- The cron queue to enqueue tasks in.
• forEachRealTld -- Specifies that the task should be run in each TLD of type REAL. This can be combined with forEachTestTld.
• forEachTestTld -- Specifies that the task should be run in each TLD of type TEST. This can be combined with forEachRealTld.
• runInEmpty -- Specifies that the task should be run globally, i.e. just once, rather than individually per TLD. This is provided to allow tasks to retry. It is called "runInEmpty" for historical reasons.
• excludes -- A list of TLDs to exclude from processing.
• jitterSeconds -- The execution of each per-TLD task is delayed by a different random number of seconds between zero and this max value.

Cloud Datastore

The Domain Registry platform uses Cloud Datastore as its primary database. Cloud Datastore is a NoSQL document database that provides automatic horizontal scaling, high performance, and high availability. All information that is persisted to Cloud Datastore takes the form of Java classes annotated with @Entity that are located in the model package. The Objectify library is used to persist instances of these classes in a format that Datastore understands.

A brief overview of the different entity types found in the App Engine Datastore Viewer may help administrators understand what they are seeing. Note that some of these entities are part of App Engine tools that are outside of the domain registry codebase:

• \_AE\_* -- These entities are created by App Engine.
• \_ah\_SESSION -- These entities track App Engine client sessions.
• \_GAE\_MR\_* -- These entities are generated by App Engine while running MapReduces.
• BackupStatus -- There should only be one of these entities, used to maintain the state of the backup process.
• Cancellation -- A cancellation is a special type of billing event which represents the cancellation of another billing event such as a OneTime or Recurring.
• ClaimsList, ClaimsListShard, and ClaimsListSingleton -- These entities store the TMCH claims list, for use in trademark processing.
• CommitLog* -- These entities store the commit log information.
• ContactResource -- These hold the ICANN contact information (but not registrar contacts, who have a separate entity type).
• Cursor -- We use Cursor entities to maintain state about daily processes, remembering which dates have been processed. For instance, for the RDE export, Cursor entities maintain the date up to which each TLD has been exported.
• DomainApplicationIndex -- These hold domain applications received during the sunrise period.
• DomainBase -- These hold the ICANN domain information.
• DomainRecord -- These are used during the DNS update process.
• EntityGroupRoot -- There is only one EntityGroupRoot entity, which serves as the Datastore parent of many other entities.
• EppResourceIndex -- These entities allow enumeration of EPP resources (such as domains, hosts and contacts), which would otherwise be difficult to do in Datastore.
• ExceptionReportEntity -- These entities are generated automatically by ECatcher, a Google-internal logging and debugging tool. Non-Google users should not encounter these entries.
• ForeignKeyContactIndex, ForeignKeyDomainIndex, and ForeignKeyHostIndex -- These act as a unique index on contacts, domains and hosts, allowing transactional lookup by foreign key.
• HistoryEntry -- A HistoryEntry is the record of a command which mutated an EPP resource. It serves as the parent of BillingEvents and PollMessages.
• HostRecord -- These are used during the DNS update process.
• HostResource -- These hold the ICANN host information.
• Lock -- Lock entities are used to control access to a shared resource such as an App Engine queue. Under ordinary circumstances, these locks will be cleaned up automatically, and should not accumulate.
• LogsExportCursor -- This is a single entity which maintains the state of log export.
• MR-* -- These entities are generated by the App Engine MapReduce library in the course of running MapReduces.
• Modification -- A Modification is a special type of billing event which represents the modification of a OneTime billing event.
• OneTime -- A OneTime is a billing event which represents a one-time charge or credit to the client (as opposed to Recurring).
• pipeline-* -- These entities are also generated by the App Engine MapReduce library.
• PollMessage -- PollMessages are generated by the system to notify registrars of asynchronous responses and status changes.
• PremiumList, PremiumListEntry, and PremiumListRevision -- The standard method for determining which domain names receive premium pricing is to maintain a static list of premium names. Each PremiumList contains some number of PremiumListRevisions, each of which in turn contains a PremiumListEntry for each premium name.
• RdeRevision -- These entities are used by the RDE subsystem in the process of generating files.
• Recurring -- A Recurring is a billing event which represents a recurring charge to the client (as opposed to OneTime).
• Registrar -- These hold information about client registrars.
• RegistrarContact -- Registrars have contacts just as domains do. These are stored in a special RegistrarContact entity.
• RegistrarCredit and RegistrarCreditBalance -- The system supports the concept of a registrar credit balance, which is a pool of credit that the registrar can use to offset amounts they owe. This might come from promotions, for instance. These entities maintain registrars' balances.
• Registry -- These hold information about the TLDs supported by the Registry system.
• RegistryCursor -- These entities are the predecessor to the Cursor entities. We are no longer using them, and will be deleting them soon.
• ReservedList -- Each ReservedList entity represents an entire list of reserved names which cannot be registered. Each TLD can have one or more attached reserved lists.
• ServerSecret -- this is a single entity containing the secret numbers used for generating tokens such as XSRF tokens.
• SignedMarkRevocationList -- The entities together contain the Signed Mark Data Revocation List file downloaded from the TMCH MarksDB each day. Each entity contains up to 10,000 rows of the file, so depending on the size of the file, there will be some handful of entities.
• TmchCrl -- This is a single entity containing ICANN's TMCH CA Certificate Revocation List.

Cloud Storage buckets

The Domain Registry platform uses Cloud Storage for bulk storage of large flat files that aren't suitable for Datastore. These files include backups, RDE exports, Datastore snapshots (for ingestion into BigQuery), and reports. Each bucket name must be unique across all of Google Cloud Storage, so we use the common recommended pattern of prefixing all buckets with the name of the App Engine app (which is itself globally unique). Most of the bucket names are configurable, but the defaults are as follows, with PROJECT standing in as a placeholder for the App Engine app name:

• PROJECT-billing -- Monthly invoice files for each registrar.
• PROJECT-commits -- Daily exports of commit logs that are needed for potentially performing a restore.
• PROJECT-domain-lists -- Daily exports of all registered domain names per TLD.
• PROJECT-gcs-logs -- This bucket is used at Google to store the GCS access logs and storage data. This bucket is not required by the Registry system, but can provide useful logging information. For instructions on setup, see the Cloud Storage documentation.
• PROJECT-icann-brda -- This bucket contains the weekly ICANN BRDA files. There is no lifecycle expiration; we keep a history of all the files. This bucket must exist for the BRDA process to function.
• PROJECT-icann-zfa -- This bucket contains the most recent ICANN ZFA files. No lifecycle is needed, because the files are overwritten each time.
• PROJECT-rde -- This bucket contains RDE exports, which should then be regularly uploaded to the escrow provider. Lifecycle is set to 90 days. The bucket must exist.
• PROJECT-reporting -- Contains monthly ICANN reporting files.
• PROJECT-snapshots -- Contains daily exports of Datastore entities of types defined in ExportConstants.java. These are imported into BigQuery daily to allow for in-depth querying.
• PROJECT.appspot.com -- Temporary MapReduce files are stored here. By default, the App Engine MapReduce library places its temporary files in a bucket named {project}.appspot.com. This bucket must exist. To keep temporary files from building up, a 90-day or 180-day lifecycle should be applied to the bucket, depending on how long you want to be able to go back and debug MapReduce problems. At 30 GB per day of generate temporary files, this bucket may be the largest consumer of storage, so only save what you actually use.

Commit logs

Web.xml

Cursors