Read the article: Architecture rookie elastic scheduling system design
Article URL: https: //mp.weixin.qq.com/s __biz = MzIzOTU0NTQ0MA == & mid = 2247487005 & idx = 1 & sn = fd7db29c85eac932f2d8c86ae6f8a017 & chksm = e9293312de5eba043eb6b6f745fedf1358779f821d34ab3385c46c4e527d0cd7053581ddca63 & scene = 21 # wechat_redirect?
In the prior scheduling elastic appear rookie overall resource usage are at a relatively low level, because:
( 1 ) single-line application typically through pressure measurement performance, and projected traffic binding empirically determine the desired number of containers. In this way a large extent be disturbed assess subjective factors in estimating traffic flow is also usually a relatively large redundancy.
( 2 ) the conventional mode, the operating frequency of a scalable application packet capacity low, which makes estimating traffic required at peak and daily traffic for some time (typically in units of months) to the development of the business as the evaluation criteria. The general appearance peak flow periods may account for only a fraction of the time of day, off-peak periods there have a lot of waste of resources.
Elastic scheduling Chui then be able to bring greater benefits to use, but it is not adapted to all companies or organizations, and the rookie was able to achieve success in landing, depending on the following reasons:
(1) operational characteristics rookie is to coordinate business, cp , information flow between consumers, and orders the transfer of logistics and more interactive features long link also determines the flow of information flow is greater than the practical operation .
(2) rookie in 2017 after the beginning of the full realization of the container and access to hybrid cloud infrastructure systems, has completed the resource management from the " machine-oriented " to " application-oriented " transformation, application deployment, scaling capacity and other core operation and maintenance process has been greatly simplified and efficiency.
(3) novice Ali Group specification compliant application, the elastic scheduler can be read directly alimonitor , Eagle Eye, alimetrics monitoring operation and maintenance data, tools, and core technology stack used by applications to obtain substantially convergence, this bit flexible scheduled landing provides adequate environmental basis.
The basic mode is scheduled elastic closed-loop feedback. This mode is advantageous in that :( 1 ) have a certain degree of self-evolution capability. ( 2 ) a higher level of abstraction to the configuration parameters of the mass, in order to solve common problems.
Ark elastic scheduled three decisions:
(1) The first layer is a policy decision, policy decision-making by a number of different strategies composed of, and support rapid expansion. Between policy logic completely isolated, are independent output operation (expansion, volume reduction, constant) and quantity calculation is done for each policy. In order to accommodate heterogeneous between different applications, each packet may also activate or deactivate a different policy based on the actual situation.
(2) a second layer polymeric decision, the decision to collect the polymerization results of all the policy decision of the first layer, and to obtain the aggregation rules based on a combined < operation number > group. This layer rule is simple: when there is simultaneous expansion and contraction capacity of decision-making results to expansion prevail, ignoring volume reduction results; when there are multiple expansion results to the largest number of expansion results as the final result; when there are multiple when a volume reduction result, a volume reduction as a result of a small number of the final result.
(3) The third layer is the implementation of the decision, the main part of this decision will take into account some of the rules, eventually told scalable content service: whether or not scalable, the number of containers to be scalable, so if you want a volume reduction volume reduction which several specific container, if it is so specific expansion of container specifications, expansion to the engine room and so on.
Decision Strategies Ark elastic support rapid scale-scheduling, now it contains more decision-making strategy, part of the policy is tested to verify the state where most of the core of a few, but also the first to run on-line strategies are introduced:
(1) Resource security policy: security policy concern is resource system resource usage.
(2) Resource Optimization Strategies: Resource optimization strategy is also concerned about the use of system resources, but it exists to reclaim resources when the system is idle. Follow the same above three system parameters, when the three simultaneously below the lower threshold, initiating volume reduction.
(3) Time Strategy: The current decision-making model is resilient posteriori, ie after the occurrence of a threshold violation, it initiated expansion.
(4) Service security policy: service security policy is the most complicated of a strategy, at present, contains a message queue service Consumer , RPC services, HTTP services three. Every day at least half of the new capacity is initiated by the service security policy.
Many flexible policy scheduling system will select services qps as the most important consideration, but we think after a series of preliminary discussions, decided to abandon qps using RT , mainly for the following reasons:
(1) qps is a service of the variables, it changes limited to the use of the user.
(2) by the current qps and can withstand the maximum single qps to get the current number of containers, completely isolated resources, and each query resource use approximately equal when established.
By reading this blog, I, burr handling of elasticity scheduling system architecture, to calculate the number of expansion and scaling a certain understanding.