/// <summary>
/// Decides whether a scaling move is allowed. At this point, we don't know if Azure Cache for Redis business
/// rules allow scaling from the current to the target size. We just decide whether it is reasonable based on
/// our knowledge of our production workload.
///
/// The autoscaler will figure out how to reach the desired plan.
/// </summary>
public static bool IsScalingAllowed(
RedisClusterSize current,
RedisClusterSize target,
ModelContext modelContext)
{
// Cluster must be able to handle the amount of data we'll give it, with some overhead in case of
// production issues. Notice we don't introduce a per-shard restriction; reason for this is that the shards
// distribute keys evenly.
if (target.ClusterMemorySizeMb < modelContext.MinimumAllowedClusterMemoryMb)
{
return(false);
}
// Cluster must be able to handle the amount of operations needed. Notice we don't introduce a per-shard
// restriction; reason for this is that the shards distribute keys evenly.
if (target.EstimatedRequestsPerSecond < modelContext.MinimumAllowedClusterRps)
{
return(false);
}
// Disallow going over the maximum allowed cluster memory
if (modelContext.MaximumAllowedClusterMemoryMb != null && target.ClusterMemorySizeMb > modelContext.MaximumAllowedClusterMemoryMb.Value)
{
return(false);
}
return(true);
}
public static bool IsDownScale(RedisClusterSize from, RedisClusterSize to)
{
// Same tier (i.e. P3) means that we only care about shards
if (from.Tier.Equals(to.Tier))
{
return(from.Shards > to.Shards);
}
// Distinct tier, but same number of shards means we only care about tier
if (from.Shards == to.Shards)
{
return(IsDownScale(from.Tier, to.Tier));
}
// Distinct tier and distinct number of shards gets a bit more complicated, so we turn to looking at memory
// capacity, server capacity, and cost
if (from.ClusterMemorySizeMb > to.ClusterMemorySizeMb)
{
return(true);
}
if (from.EstimatedRequestsPerSecond > to.EstimatedRequestsPerSecond)
{
return(true);
}
if (from.MonthlyCostUsd > to.MonthlyCostUsd)
{
return(true);
}
return(false);
}
public ModelOutput(RedisClusterSize targetClusterSize, ModelContext modelContext, double cost, IReadOnlyList <RedisClusterSize> scalePath)
{
TargetClusterSize = targetClusterSize;
ModelContext = modelContext;
Cost = cost;
ScalePath = scalePath;
}
private async Task <BoolResult> SubmitScaleRequestAsync(RedisClusterSize targetClusterSize, CancellationToken cancellationToken)
{
var instance = RedisCache.Update();
if (!ClusterSize.Tier.Equals(targetClusterSize.Tier))
{
switch (targetClusterSize.Tier.Plan)
{
case RedisPlan.Basic:
instance = instance.WithBasicSku(targetClusterSize.Tier.Capacity);
break;
case RedisPlan.Standard:
instance = instance.WithStandardSku(targetClusterSize.Tier.Capacity);
break;
case RedisPlan.Premium:
instance = instance.WithPremiumSku(targetClusterSize.Tier.Capacity);
break;
}
}
if (ClusterSize.Shards != targetClusterSize.Shards)
{
instance = instance.WithShardCount(targetClusterSize.Shards);
}
await instance.ApplyAsync(cancellationToken);
return(BoolResult.Success);
}
public static Result <ModelOutput> Predict(RedisClusterSize currentClusterSize, ModelContext modelContext)
{
var shortestPaths = ComputeAllowedPaths(currentClusterSize, modelContext);
var eligibleClusterSizes = shortestPaths
.Select(kvp => (Size: kvp.Key, Node: kvp.Value))
// Find all plans that we can reach from the current one via scaling operations, and that we allow scaling to
.Where(entry => entry.Node.ShortestDistanceFromSource != double.PositiveInfinity && IsScalingAllowed(currentClusterSize, entry.Size, modelContext))
// Compute the cost of taking the given route
.Select(entry => (entry.Size, entry.Node, Cost: CostFunction(currentClusterSize, entry.Size, modelContext, shortestPaths)))
.ToList();
// Rank them by cost ascending
var costSorted = eligibleClusterSizes
.OrderBy(pair => pair.Cost)
.ToList();
if (costSorted.Count == 0)
{
return(new Result <ModelOutput>(errorMessage: "No cluster size available for scaling"));
}
return(new ModelOutput(
targetClusterSize: costSorted[0].Size,
modelContext: modelContext,
cost: costSorted[0].Cost,
scalePath: RedisScalingUtilities.ComputeShortestPath(shortestPaths, currentClusterSize, costSorted[0].Size)));
}
private RedisInstance(IAzure azure, string resourceId, IRedisCache redisCache, RedisClusterSize clusterSize)
{
Contract.RequiresNotNullOrEmpty(resourceId);
_azure = azure;
_resourceId = resourceId;
RedisCache = redisCache;
ClusterSize = clusterSize;
}
public async Task <BoolResult> RefreshAsync(CancellationToken cancellationToken = default)
{
return((await GenerateInstanceMetadataAsync(_azure, _resourceId, cancellationToken))
.Select(result =>
{
RedisCache = result.Cache;
ClusterSize = result.Size;
return BoolResult.Success;
}));
}
public static async Task <Result <(IRedisCache Cache, RedisClusterSize Size)> > GenerateInstanceMetadataAsync(IAzure azure, string resourceId, CancellationToken cancellationToken = default)
{
// TODO: error handling
var redisCache = await azure.RedisCaches.GetByIdAsync(resourceId, cancellationToken);
var clusterSize = RedisClusterSize.FromAzureCache(redisCache).ThrowIfFailure();
Contract.AssertNotNull(clusterSize);
return(new Result <(IRedisCache Cache, RedisClusterSize Size)>((redisCache, clusterSize)));
}
private Task <BoolResult> RequestScaleAsync(OperationContext context, RedisClusterSize targetClusterSize)
{
string extraMessage = $"CurrentClusterSize=[{ClusterSize}] TargetClusterSize=[{targetClusterSize}]";
return(context.PerformOperationAsync(Tracer, async() =>
{
if (ClusterSize.Equals(targetClusterSize))
{
return new BoolResult(errorMessage: $"No-op scale request attempted (`{ClusterSize}` -> `{targetClusterSize}`) on instance `{Name}`");
}
if (!RedisScalingUtilities.CanScale(ClusterSize, targetClusterSize))
{
return new BoolResult(errorMessage: $"Scale request `{ClusterSize}` -> `{targetClusterSize}` on instance `{Name}` is disallowed by Azure Cache for Redis");
}
if (!IsReadyToScale)
{
return new BoolResult(errorMessage: $"Redis instance `{Name}` is not ready to scale, current provisioning state is `{RedisCache.ProvisioningState}`");
}
var instance = RedisCache.Update();
if (!ClusterSize.Tier.Equals(targetClusterSize.Tier))
{
switch (targetClusterSize.Tier.Plan)
{
case RedisPlan.Basic:
instance = instance.WithBasicSku(targetClusterSize.Tier.Capacity);
break;
case RedisPlan.Standard:
instance = instance.WithStandardSku(targetClusterSize.Tier.Capacity);
break;
case RedisPlan.Premium:
instance = instance.WithPremiumSku(targetClusterSize.Tier.Capacity);
break;
}
}
if (ClusterSize.Shards != targetClusterSize.Shards)
{
instance = instance.WithShardCount(targetClusterSize.Shards);
}
await instance.ApplyAsync(context.Token);
return BoolResult.Success;
},
extraStartMessage: extraMessage,
extraEndMessage: _ => extraMessage,
pendingOperationTracingInterval: TimeSpan.FromMinutes(1)));
}
public static IReadOnlyList <RedisClusterSize> ComputeShortestPath(RedisClusterSize from, RedisClusterSize to, Func <RedisClusterSize, IEnumerable <RedisClusterSize> > neighbors, Func <RedisClusterSize, RedisClusterSize, double> weight, IReadOnlyList <RedisClusterSize>?vertices = null)
{
if (from.Equals(to))
{
return(Array.Empty <RedisClusterSize>());
}
vertices ??= RedisClusterSize.Instances;
var shortestPaths = ComputeOneToAllShortestPath(vertices, neighbors, weight, from);
return(ComputeShortestPath(shortestPaths, from, to));
}
/// <summary>
/// This function embodies the concept of "how much does it cost to switch from
/// <paramref name="current"/> to <paramref name="target"/>". At this point, we can assume that:
/// - The two input sizes are valid states to be in
/// - We can reach the target from current via some amount of autoscaling operations
/// Hence, we're just ranking amonst the many potential states.
/// </summary>
public static double CostFunction(RedisClusterSize current, RedisClusterSize target, ModelContext modelContext, IReadOnlyDictionary <RedisClusterSize, RedisScalingUtilities.Node> shortestPaths)
{
// Switching to the same size (i.e. no op) is free
if (current.Equals(target))
{
return(0);
}
var shortestPath = RedisScalingUtilities.ComputeShortestPath(shortestPaths, current, target);
Contract.Assert(shortestPath.Count > 0);
// Positive if we are spending more money, negative if we are saving
return((double)(target.MonthlyCostUsd - current.MonthlyCostUsd));
}
public static Result <IRedisInstance> FromPreloaded(IAzure azure, IRedisCache redisCache, bool readOnly)
{
return(RedisClusterSize
.FromAzureCache(redisCache)
.Select(clusterSize =>
{
if (readOnly)
{
return (IRedisInstance) new ReadOnlyRedisInstance(azure, redisCache.Id, redisCache, clusterSize);
}
else
{
return (IRedisInstance) new RedisInstance(azure, redisCache.Id, redisCache, clusterSize);
}
}));
}
private async Task <BoolResult> RequestScaleAsync(RedisClusterSize targetClusterSize, CancellationToken cancellationToken = default)
{
if (ClusterSize.Equals(targetClusterSize))
{
return(new BoolResult(errorMessage: $"No-op scale request attempted (`{ClusterSize}` -> `{targetClusterSize}`) on instance `{Name}`"));
}
if (!RedisScalingUtilities.CanScale(ClusterSize, targetClusterSize))
{
return(new BoolResult(errorMessage: $"Scale request `{ClusterSize}` -> `{targetClusterSize}` on instance `{Name}` is disallowed by Azure Cache for Redis"));
}
if (!IsReadyToScale)
{
return(new BoolResult(errorMessage: $"Redis instance `{Name}` is not ready to scale, current provisioning state is `{RedisCache.ProvisioningState}`"));
}
var instance = RedisCache.Update();
if (!ClusterSize.Tier.Equals(targetClusterSize.Tier))
{
switch (targetClusterSize.Tier.Plan)
{
case RedisPlan.Basic:
instance = instance.WithBasicSku(targetClusterSize.Tier.Capacity);
break;
case RedisPlan.Standard:
instance = instance.WithStandardSku(targetClusterSize.Tier.Capacity);
break;
case RedisPlan.Premium:
instance = instance.WithPremiumSku(targetClusterSize.Tier.Capacity);
break;
}
}
if (ClusterSize.Shards != targetClusterSize.Shards)
{
instance = instance.WithShardCount(targetClusterSize.Shards);
}
await instance.ApplyAsync(cancellationToken);
return(BoolResult.Success);
}
private async Task <BoolResult> RequestScaleAsync(RedisClusterSize targetClusterSize, CancellationToken cancellationToken = default)
{
if (ClusterSize.Equals(targetClusterSize))
{
return(new BoolResult(errorMessage: $"No-op scale request attempted (`{ClusterSize}` -> `{targetClusterSize}`) on instance `{Name}`"));
}
if (!RedisScalingUtilities.CanScale(ClusterSize, targetClusterSize))
{
return(new BoolResult(errorMessage: $"Scale request `{ClusterSize}` -> `{targetClusterSize}` on instance `{Name}` is disallowed by Azure Cache for Redis"));
}
if (!IsReadyToScale)
{
return(new BoolResult(errorMessage: $"Redis instance `{Name}` is not ready to scale, current provisioning state is `{RedisCache.ProvisioningState}`"));
}
return(await SubmitScaleRequestAsync(targetClusterSize, cancellationToken));
}
public static bool CanScale(RedisClusterSize from, RedisClusterSize to)
{
if (from.Equals(to))
{
return(true);
}
if (!CanScale(from.Tier, to.Tier))
{
return(false);
}
if (from.Shards != to.Shards && !from.Tier.Equals(to.Tier))
{
// Azure can't change both shards and tiers at once, we need to do them one at a time.
return(false);
}
return(true);
}
public static TimeSpan ExpectedScalingDelay(RedisClusterSize from, RedisClusterSize to)
{
Contract.Requires(CanScale(from, to));
if (from.Equals(to))
{
return(TimeSpan.Zero);
}
if (from.Tier.Equals(to.Tier))
{
// The tier is the same, so autoscaling will be either adding or reducing shards
var shardDelta = Math.Abs(from.Shards - to.Shards);
return(TimeSpan.FromTicks(Constants.RedisScaleTimePerShard.Ticks * shardDelta));
}
else
{
// Tier changed, which means the number of shards didn't. However, we will take the same amount of time
// as the amount of shards that need to change tier.
Contract.Assert(from.Shards == to.Shards);
return(TimeSpan.FromTicks(Constants.RedisScaleTimePerShard.Ticks * from.Shards));
}
}
public static Result <RedisInstance> FromPreloaded(IAzure azure, IRedisCache redisCache)
{
return(RedisClusterSize
.FromAzureCache(redisCache)
.Select(clusterSize => new RedisInstance(azure, redisCache.Id, redisCache, clusterSize)));
}
internal RedisInstance(IAzure azure, string resourceId, IRedisCache redisCache, RedisClusterSize clusterSize)
: base(azure, resourceId, redisCache, clusterSize)
{
}
public Node(RedisClusterSize clusterSize)
{
ClusterSize = clusterSize;
}
public static Dictionary <RedisClusterSize, Node> ComputeOneToAllShortestPath(IReadOnlyList <RedisClusterSize> vertices, Func <RedisClusterSize, IEnumerable <RedisClusterSize> > neighbors, Func <RedisClusterSize, RedisClusterSize, double> weight, RedisClusterSize from)
{
// We need to find a valid scale order to reach the target cluster size from the current one. To find it,
// create an implicit graph G = (V, E) where V is the set of Redis sizes, and E is the set of valid
// scalings given by the CanScale relation. In this graph, finding a shortest path between the current and
// target sizes is equivalent to figuring out a way to scale among them optimally, as given by whatever
// weight function we choose.
var translation = new Dictionary <RedisClusterSize, Node>(capacity: vertices.Count);
var minPriorityQueue = new SortedSet <Node>(comparer: NodeComparer.Instance);
foreach (var vertex in vertices)
{
var node = new Node(vertex);
if (vertex.Equals(from))
{
node.ShortestDistanceFromSource = 0;
}
minPriorityQueue.Add(node);
translation[vertex] = node;
}
while (minPriorityQueue.Count > 0)
{
var node = minPriorityQueue.Min;
Contract.AssertNotNull(node);
minPriorityQueue.Remove(node);
if (node.Visited)
{
continue;
}
node.Visited = true;
foreach (var target in neighbors(node.ClusterSize))
{
var adjacent = translation[target];
Contract.AssertNotNull(adjacent);
var distanceThroughNode = node.ShortestDistanceFromSource + weight(node.ClusterSize, target);
if (distanceThroughNode >= adjacent.ShortestDistanceFromSource)
{
continue;
}
// Typically, we'd like to do a decrease priority operation here. This is a work-around to avoid
// using a more complex data structure.
minPriorityQueue.Remove(adjacent);
adjacent.ShortestDistanceFromSource = distanceThroughNode;
adjacent.Predecessor = node;
minPriorityQueue.Add(adjacent);
}
}
return(translation);
}
public static IReadOnlyList <RedisClusterSize> ComputeShortestPath(IReadOnlyDictionary <RedisClusterSize, Node> shortestPaths, RedisClusterSize from, RedisClusterSize to)
{
if (from.Equals(to))
{
return(Array.Empty <RedisClusterSize>());
}
var path = new List <Node>();
var current = shortestPaths[to];
while (current.Predecessor != null)
{
path.Add(current);
current = current.Predecessor;
}
if (!current.ClusterSize.Equals(from))
{
return(Array.Empty <RedisClusterSize>());
}
path.Reverse();
return(path.Select(p => p.ClusterSize).ToList());
}
private async Task ComputeMemoryRelatedFeaturesAsync(DateTime now, string redisAzureId, RedisClusterSize currentClusterSize, ModelContext modelContext, CancellationToken cancellationToken = default)
{
var startTimeUtc = now - _configuration.UsedMemoryLookback;
var endTimeUtc = now;
// This maximum is the maximum across all shards. There's no way to tell what the memory usage is for
// everything precisely without fetching metrics for each shard individually.
var usedMemoryTasks =
Enumerable.Range(0, 10)
.Select(shard =>
_monitorManagementClient.GetMetricAsync(
redisAzureId,
AzureRedisShardMetric.UsedMemory.ToMetricName(shard: shard),
startTimeUtc,
endTimeUtc,
_configuration.UsedMemoryAggregationInterval,
AggregationType.Maximum,
cancellationToken));
var usedMemoryBytes = await Task.WhenAll(usedMemoryTasks);
var groupedMetrics = usedMemoryBytes
.SelectMany(measurements => measurements.Select((measurement, index) => (measurement, index)))
.GroupBy(entry => entry.index)
.OrderBy(group => group.Key)
.Select(group => group.Sum(entry => entry.measurement.Value ?? 0))
.ToList();
// Metric is reported in bytes, we use megabytes for everything
var expectedClusterMemoryUsageMb = groupedMetrics.Max() / 1e+6;
modelContext.MinimumAllowedClusterMemoryMb = (1 + _configuration.MinimumExtraMemoryAvailable) * expectedClusterMemoryUsageMb;
modelContext.MaximumAllowedClusterMemoryMb = _configuration.MaximumClusterMemoryAllowedMb;
}
private async Task ComputeWorkloadRelatedFeaturesAsync(DateTime now, string redisAzureId, RedisClusterSize currentClusterSize, ModelContext modelContext, CancellationToken cancellationToken)
{
var startTimeUtc = now - _configuration.WorkloadLookback;
var endTimeUtc = now;
// This maximum is the maximum across all shards. There's no way to tell what the memory usage is for
// everything precisely without fetching metrics for each shard individually.
var operationsPerSecondTasks =
Enumerable.Range(0, 10)
.Select(shard =>
_monitorManagementClient.GetMetricAsync(
redisAzureId,
AzureRedisShardMetric.OperationsPerSecond.ToMetricName(shard: shard),
startTimeUtc,
endTimeUtc,
_configuration.WorkloadAggregationInterval,
AggregationType.Maximum,
cancellationToken));
var operationsPerSecond = await Task.WhenAll(operationsPerSecondTasks);
var groupedMetrics = operationsPerSecond
.SelectMany(measurements => measurements.Select((measurement, index) => (measurement, index)))
.GroupBy(entry => entry.index)
.OrderBy(group => group.Key)
.Select(group => group.Sum(entry => entry.measurement.Value ?? 0))
.ToList();
// ops/s scales linearly with shards
var expectedClusterRps = groupedMetrics.Max();
modelContext.MinimumAllowedClusterRps = (1 + _configuration.MinimumWorkloadExtraPct) * expectedClusterRps;
}
private async Task <ModelContext> ComputeFeaturesAsync(DateTime now, string redisAzureId, RedisClusterSize currentClusterSize, CancellationToken cancellationToken = default)
{
var modelContext = new ModelContext();
await ComputeMemoryRelatedFeaturesAsync(now, redisAzureId, currentClusterSize, modelContext, cancellationToken);
await ComputeWorkloadRelatedFeaturesAsync(now, redisAzureId, currentClusterSize, modelContext, cancellationToken);
return(modelContext);
}
public static IReadOnlyDictionary <RedisClusterSize, RedisScalingUtilities.Node> ComputeAllowedPaths(RedisClusterSize currentClusterSize, ModelContext modelContext)
{
// We need to reach the target cluster size, but we can't do it in one shot because business rules won't
// let us, so we need to compute a path to get to it. This is probably the most complex part of the
// algorithm, there are several competing aspects we want to optimize for, in descending importance:
// - We want for memory to get to the target level ASAP
// - We want to keep the number of shards as stable as possible, given that changing them can cause build
// failures
// - We'd like to get there in the fewest amount of time possible
// - The route needs to be deterministic, so that if we are forced to stop and re-compute it we'll take
// the same route.
// - We'd like to minimize the cost of the route
// Multi-constraint optimization over graphs is NP-complete and algorithms are hard to come up with, so we
// do our best.
Func <RedisClusterSize, IEnumerable <RedisClusterSize> > neighbors =
currentClusterSize => currentClusterSize.ScaleEligibleSizes.Where(targetClusterSize =>
{
// Constrain paths to downscale at most one shard at the time. This only makes paths longer, so it
// is safe. The reason behind this is that the service doesn't really tolerate big reductions.
if (targetClusterSize.Shards < currentClusterSize.Shards)
{
return(targetClusterSize.Shards == currentClusterSize.Shards - 1);
}
return(true);
});
Func <RedisClusterSize, RedisClusterSize, double> weight =
(from, to) =>
{
// This factor is used to avoid transitioning to any kind of intermediate plan that may cause a
// production outage. If we don't have it, we may transition into a state in which we have less
// cluster memory available than we need. By adjusting the weight function, we guarantee that
// this only happens iff there is no better path; moreover, we will always choose the lesser of
// two evils if given no choice.
double clusterMemoryPenalization = 0;
var delta = to.ClusterMemorySizeMb - modelContext.MinimumAllowedClusterMemoryMb;
if (delta < 0)
{
// The amount of cluster memory is less than we need, so we penalize taking this path by
// adding the amount of memory that keeps us away from the target.
clusterMemoryPenalization = -delta;
}
// This needs to be at least one so we don't pick minimum paths that are arbitrarily long
return(1 + clusterMemoryPenalization);
};
return(RedisScalingUtilities.ComputeOneToAllShortestPath(vertices: RedisClusterSize.Instances, neighbors: neighbors, weight: weight, from: currentClusterSize));
}
