Build HA services using synchronously replicated disks


Regional Persistent Disk and Hyperdisk Balanced High Availability are storage options that let you implement high availability (HA) services in Compute Engine. Regional Persistent Disk and Hyperdisk Balanced High Availability synchronously replicate data between two zones in the same region and ensure HA for disk data for up to one zonal failure.

Building high availability services with replicated disks

This section explains how you can build HA services with regional Persistent Disk or Hyperdisk Balanced High Availability disks.

Design considerations

Before you start designing a HA service, understand the characteristics of the application, the file system, and the operating system. These characteristics are the basis for the design and can rule out various approaches. For example, if an application does not support application-level replication, some corresponding design options are not applicable.

Similarly, if the application, the file system, or the operating system are not crash tolerant, then using regional Persistent Disk or Hyperdisk Balanced High Availability (Preview) disks, or even zonal disk snapshots, might not be an option. Crash tolerance is defined as the ability to recover from an abrupt termination without losing or corrupting data that was already committed to a disk prior to the crash.

Consider the following when designing for high availability:

  • The effect on the application of using Hyperdisk Balanced High Availability, regional Persistent Disk, or other solutions
  • Disk write performance
  • The service recovery time objective - how quickly your service must recover from a zonal outage and the SLA requirements
  • The cost to build a resilient and reliable service architecture

In terms of cost, use the following options for synchronous and asynchronous application replication:

  • Use two instances of the database and VM. In this case the following items determine the total cost:

    • VM instance costs
    • Persistent Disk or Hyperdisk costs
    • Costs of maintaining application replication
  • Use a single VM with synchronously replicated disks. To achieve high availability with a regional Persistent Disk or Hyperdisk Balanced High Availability disk, use the same VM instance and disk components as the previous option, but also include a synchronously replicated disk. Regional Persistent Disks and Hyperdisk Balanced High Availability disks are double the cost per byte compared to zonal disks because they are replicated in two zones.

    However, using synchronously replicated disks might reduce your maintenance cost because the data is automatically written to two replicas without the requirement of maintaining application replication.

  • Don't start the secondary VM until failover is required. You can reduce host costs even more by starting the secondary VM only on demand during failover rather than maintaining the VM as an active standby VM.

Compare cost, performance, and resiliency

The following table highlights the trade-offs in cost, performance, and resiliency for the different service architectures.

HA service
architecture
Zonal disk
snapshots
Application level
synchronous
Application level
asynchronous
Synchronously replicated disks
Protects against application, VM, zone failure1
Mitigation against application corruption (Example: application crash-intolerance) 2 2
Cost $ $$

  • Two instances of the database or VM running, application replication setup and maintenance, cross zone networking.
$$

  • Two instances of the database or VM running, application replication setup and maintenance, cross zone networking.
$1.5x - $$

  • The costs are the same as application replication if you use an active standby VM. You can achieve a lower cost by spinning up a backup VM on demand during failover. There is no charge for cross zone networking between disk replicas.
Application performance

  • No effect


  • Trade-off on application performance with synchronous replication


  • No effect


  • No effect for most applications
Suited for application with low RPO requirement (Very low tolerance for data loss)

  • Data loss depending on when the snapshot was taken


  • No data loss3


  • Data loss because replication is asynchronous


  • No data loss
Storage recovery time from disaster4
  • O(minutes)
  • O(seconds)
  • O(seconds)
  • O(seconds)—to force attach the disk to a standby VM instance

1 Using replicated disks or snapshots is not sufficient to protect from and mitigate against failures and corruptions. Your application, file system, and possibly other software components must be crash consistent or use some sort of quiescing.

2 The replication of some applications provides mitigation against some application corruptions. For example, MySQL primary application corruption doesn't cause its replica VM instances to become corrupted as well. Review your application's documentation for details.

3 Data loss means unrecoverable loss of data committed to persistent storage. Any non-committed data is still lost.

4 Failover performance doesn't include file system check and application recovery and load after failover.

Building HA database services using synchronously replicated disks

This section covers high level concepts for building HA solutions for stateful database services (MySQL, Postgres, etc.) using Compute Engine with Regional Persistent Disks and Hyperdisk Balanced High Availability disks (Preview).

If there are broad outages in Google Cloud, for example, if a whole region becomes unavailable, your application might become unavailable. Depending on your needs, consider cross-regional replication techniques or Persistent Disk Asynchronous Replication (PD Async Replication) for even higher availability.

Database HA configurations typically have at least two VM instances. Preferably these VM instances are part of one or more managed instance groups:

  • A primary VM instance in the primary zone
  • A standby VM instance in a secondary zone

A primary VM instance has at least two disks: a boot disk, and a replicated disk. The replicated disk contains database data and any other mutable data that should be preserved to another zone in case of an outage.

A standby VM instance requires a separate boot disk to be able to recover from configuration-related outages, which could result from an operating system upgrade, for example. Also, you can't force attach a boot disk to another VM during a failover.

The primary and standby VM instances are configured to use a load balancer with the traffic directed to the primary VM based on health check signals. The disaster recovery scenario for data outlines other failover configurations, which might be more appropriate for your scenario.

Challenges with database replication

The following table lists some common challenges with setting up and managing application synchronous or semi-synchronous replication (like MySQL) and how they compare to synchronous disk replication with Regional Persistent Disk and Hyperdisk Balanced High Availability disks (Preview).

Challenges Application synchronous
or semi-synchronous replication
Synchronous disk replication
Maintaining stable replication between primary and failover replica. There are a number of things that could go wrong and cause a VM instance to fall out of HA mode:
  1. Misconfiguration of replication parameters such as SSL certificate mismatch or missing ACL on the primary side.
  2. High load on primary VM instance causing failover replica to be unable to keep up.
  3. Bugs causing replication issues such as application issues, OS misconfiguration, or Docker failure.
  4. Infrastructure failures like CPU contention, frozen VM, or intermediate network interruption.

Storage failures are handled by Regional Persistent Disk and Hyperdisk Balanced High Availability disks. This happens transparently to the application except for a possible fluctuation in the disk's performance.

There must be user-defined health checks to reveal any application or VM issues and trigger failover.

The end-to-end failover time is longer than expected. The time taken for the failover operation doesn't have an upper bound. Waiting for all transactions to be replayed (step 2 above) could take an arbitrarily long time, depending on the schema and the load on the database.

Regional Persistent Disk and Hyperdisk Balanced High Availability disks provide synchronous replication, so the failover time is bounded by the sum of the following latencies:

  1. Creation of a secondary VM, unless there is already a standby VM instance available.
  2. Force attaching a replicated disk.
  3. Application initialization.
Split-brain To avoid split-brain, both approaches require provisions to ensure that there is only one primary at a time.

Sequence of read and write operations to disks

In determining the read and write sequences, or the order in which data is read from and written to disk, the majority of the work is done by the disk driver in your VM. As a user, you don't have to deal with the replication semantics, and can interact with the file system as usual. The underlying driver handles the sequence for reading and writing.

By default, a Compute Engine VM with Regional Persistent Disk or Hyperdisk Balanced High Availability (Preview) operates in full replication mode, where requests to read or write from disk are sent to both replicas.

In full replication mode, the following occurs:

  • When writing, a write request tries to write to both replicas and acknowledges when both writes succeed.
  • When reading, the VM sends a read request to both replicas, and returns the results from the one that succeeds. If the read request times out, another read request is sent.

If a replica falls behind or fails to acknowledge that the read or write requests completed, then the replica state is updated.

Health checks

The health checks used by the load balancer are implemented by the health check agent. The health check agent serves two purposes:

  1. The health check agent resides within the primary and secondary VMs to monitor the VM instances and communicate with the load balancer to direct traffic. This works best when configured with instance groups.
  2. The health check agent syncs with the application-specific regional control plane and makes failover decisions based on control plane behavior. The control plane must be in a zone that differs from the VM instance whose health it is monitoring.

The health check agent itself must be fault tolerant. For example, notice that, in the image that follows, the control plane is separated from the primary VM instance, which resides in zone us-central1-a, while the standby VM resides in zone us-central1-f.

Health check agent role in
         the VM.

The health check agent's role in primary and standby VM instances.

What's next