Amazon Technologies, Inc.

Bjorn Patrick Swift

Wei Xiao

Kiran-Kumar Muniswamy-Reddy

Miguel Mascarenhas Filipe

Stuart Henry Seelye Marshall

Stefano Stefani

Yijun Lu

James Robert Hamilton

Timothy Andrew Rath

Rande Anthony Blackman

David Alan Lutz

David Craig Yanacek

Maximiliano Maccanti

Stuart Henry Marshall

Methods and apparatus for token-based admission control for replicated writes are disclosed. Data objects are divided into partitions, and corresponding to each partition, at least a master replica and a slave replica are stored. A determination as to whether to accept a write request directed to the partition is made based at least in part on one or more of (a) available throughput capacity at the master replica, and (b) an indication, obtained using a token-based protocol, of available throughput capacity at the slave replica. If the write request is accepted, one or more data modification operations are initiated.

Token-based admission control for replicated writes

Providers of web services and other types of software as a service may be subject to service-level agreements requiring that response times be within a defined range. For efficiency, multiple services may be hosted on the same set of computing nodes, which may jeopardize adherence to service-level agreements. A control system may involve classifying service requests and determining desired values for measurements such as latency. An error value may be calculated based on the difference between measured and desired values. A controller may adjust a rate of capacity utilization for the computing nodes based on the current error, a history of past errors, and a prediction of future errors.

Real-time monitoring of IO load and latency

Distributed database management systems may perform range queries over the leading portion of a primary key. Non-random distribution of data may improve performance related to the processing of range queries, but may tend to cause workload to be concentrated on particular partitions. Groups of partitions may be expanded and collapsed based on detection of disproportionate workload. Disproportionate write workload may be distributed among a group of partitions that can subsequently be queried using a federated approach. Disproportionate read workload may be distributed among a group of read-only replicated partitions.

Range query capacity allocation

A pricing policy to be applied to token population changes at a token bucket used for admission control during burst-mode operations at a work target is determined. Over a time period, changes to the token population of that bucket are recorded. An amount to be charged to a client is determined, based on the recorded changes in token population and an associated pricing amount indicated in the policy.

Token-based policies burst-mode operations

Techniques are disclosed for splitting a database partition into two partitions. In embodiments, where the partition is a hash partition, the partition is split at its logical midpoint, so that half of the partition's rows are placed in each split partition. Where the partition is a hash-range partition, a number of rows of the partition are sampled. Where enough samples fall on each side of the logical midpoint, the logical midpoint is again used as the split point. Where not enough samples fall on one side of the logical midpoint, then the median of the samples is used as the split point.

Splitting database partitions

Methods and apparatus for equitable distribution of excess shared-resource throughput capacity are disclosed. A first and a second work target are configured to access a shared resource to implement accepted work requests. Admission control is managed at the work targets using respective token buckets. A first metric indicative of the work request arrival rates at the work targets during a time interval, and a second metric associated with the provisioned capacities of the work targets are determined. A number of tokens determined based on a throughput limit of the shared resource is distributed among the work targets to be used for admission control during a subsequent time interval. The number of tokens distributed to each work target is based on the first metric and/or the second metric.

Equitable distribution of excess shared-resource throughput capacity

Methods and apparatus for burst-mode admission control using token buckets are disclosed. A work request (such as a read or a write) directed to a work target is received. Based on a first criterion, a determination is made that the work target is in a burst mode of operation. A token population of a burst-mode token bucket is determined, and if the population meets a second criterion, the work request is accepted for execution.

Burst-mode admission control using token buckets

A system that implements a data storage service may store data in multiple replicated partitions on respective storage nodes. The selection of the storage nodes (or storage devices thereof) on which to store the partition replicas may be performed by administrative components that are responsible for partition management and resource allocation for respective groups of storage nodes (e.g., based on a global view of resource capacity or usage), or the selection of particular storage devices of a storage node may be determined by the storage node itself (e.g., based on a local view of resource capacity or usage). Placement policies applied at the administrative layer or storage layer may be based on the percentage or amount of provisioned, reserved, or available storage or IOPS capacity on each storage device, and particular placements (or subsequent operations to move partition replicas) may result in an overall resource utilization that is well balanced.

Storage device selection for database partition replicas

Methods and apparatus for compound token buckets usable for burst-mode admission control are disclosed. A peak burst rate and a sustained burst rate of work requests that are to be supported at a work target are determined. The maximum token populations of a peak-burst token bucket and a sustained-burst token bucket are configured, based on the peak burst rate and the sustained burst rate respectively. In response to receiving a work request directed at the work target, a determination to accept the work request for execution is made based at least in part on the token population of the peak-burst token bucket and/or the sustained-burst token bucket.

Compound token buckets for burst-mode admission control

A database service may maintain tables on behalf of clients and may provision throughput capacity for those tables. A table may be divided into multiple partitions, according to hash of the primary key values for each of the items in the table, and the items in the table may be accessed using the hash of their primary key values. Provisioned throughput capacity for the table may be divided between the partitions and used in servicing requests directed to items in the table. The service (or underlying system) may provide mechanisms for generating skew-related metrics or reports and presenting them to clients via a graphical user interface (GUI). The metrics and reports may indicate the amount of uniformity or skew in the distribution of requests across the key space for the table using histograms, heat maps, or other representations. Clients may initiate actions to correct any skewing via the GUI.

Database system providing skew metrics across a key space

A database management system may be operated by a third-party provider that hosts the system in a datacenter and provides access to the system to end users on behalf of various entities. Limits on total capacity consumption may be imposed, but may result in service outages when capacity consumption exceeds those limits. Requests to perform operations on the system may be classified. The request classifications may be associated with policies for admitting or rejecting the request. One or more token buckets representative of capacity available to the request to perform the operation may be used to determine to admit the request and updated based on the cost of performing the operation.

Growing community of inventors

Seattle, WA, United States of America

Bjorn Patrick Swift