Microservices architecture pain points

Business is concerned with communication between services

The service iteration speed slows down

Copy the code

Microservices Architecture 1.0


One gateway layer, multiple business logic layers, multiple data access layers, multiple DB/Cache, registry, configuration center

Copy the code

Microservices 2.0 Architecture - Service Grid


Infrastructure upgrading is difficult

Impact the ability and speed of delivery of infrastructure teams

Because the application introduces communication components through JAR packages

The communication component upgrade requires the application and JAR package version upgrade

Copy the code

‘Communication’ between multiple programming languages

Business one set of infrastructure per language costs a lot

Copy the code

Microservices architecture evolution


Service Grid Definition


Service Grid Architecture


Problems - Links get longer

The average RT response latency in performance will be higher

However, the loss of sidecar between native applications to the native will not exceed 1 millisecond

Copy the code

Open source framework

Earliest version linkerd


Applications can communicate with Sidecar over TCP or HTTP1.1. The two need to maintain a long connection

Copy the code

istio


  • Pilot
The control center

1. Control communication between proxies

2. Load balancing

Copy the code
  • Mixer
Data Collection Services:



After communication between proxies, some Mertics information (time consuming and number of requests) should be reported.

Synchronous reporting of all

Centralism is unreliable

Its performance affects the performance of the proxy itself

Copy the code
  • Citadel
Do authentication security related

Proxy permission authentication, for example, TLS and SSL

Copy the code

sofa mesh

Ant Financial is open source

Copy the code

architecture


1. Rewrite proxy in ISTIO

Isotio proxy is written in c++

Sofa is rewritten in go

2. Istio data collection nodes are centralized and SOFA is distributed, that is, there is a Mixer in each proxy

3. No company has suggested using ISTIO on a large scale with an inactive community

Copy the code

Sina weibo mesh


What does the service grid do


How to choice


1. The cost of business upgrade is too high. To reduce the cost of business upgrade to 0, it is necessary to be compatible with all RPC usage

2. The expectation is that the business only needs to replace the RPC JAR package with the RPC Mesh JAR package

Copy the code

Since the research train of thought


1. Be compatible with traditional physical machines and virtual machines and the cloud

2. Control center includes service management platform and data collection center

Copy the code

Architecture design


1. Data Collection Center:

A. Metric: collection time and response time

B. Trace: Distributed request tracing system APM

C, Alarm: Alarm function

2, Protocol

A. RPC: Compatible with old RPC protocols

B. Mesh includes communication protocols (HTTP1.1 and 2.0) and data protocols (Protobuff)



(Note: HTTP1.0 does not support short connections; Http1.1 and http2.0 support keep Alive long connections; TPC is a long connection; The connection is still short on the server and can be pushed to the client directly.



2. Health checks between sidecars do not pass through the registry but themselves





Copy the code

The overall process


The user initiates a command to fuse service B


1. The Service management platform, control center and data collection center are all ready-made services (introduced in previous articles), so the self-developed Service Mesh only needs to implement proxy

Service and Proxy are one process

Now you need to change it to two separate processes

3. Place both in the same POD

Copy the code

Does it matter if Sidecar dies?

No impact.

Copy the code

How to handle the failure of sidecar pod?

If the sidecar fails, it will be detected and kill the current pod. K8s will automatically restart a pod

Copy the code

What is the architecture of two applications on the same physical machine?


drift

1. Log drift



Service 1 generates log 1 on server 1

If service 1 on server 1 hangs

Start service 2 on server 2. Log 2 is generated

If log 1 and log 2 have strong dependencies

Then service 1 must be started on server 1 to continue generating logs based on log 1



2. Retry failover

If the POD hangs and restarts again, the IP will change

It doesn't matter if you try drifting to any node on the cloud

Copy the code

Complete flow chart

This complete flow chart covers

DNS, CDN, Nginx, FastDFS (or Ceph),

LVS, ServiceMash, data collection center,

Registry, Control Center, Gateway, business logic layer,

Data access layer, storage layer and other data interaction processes



It's worth a fortune if you want it

You can add me to wechat 15900411193

Copy the code

Call link


1. The purpose of protocol resolution is to be compatible with old protocols

After the client sends a request, the client service and the server service perform protocol resolution

If both mesh protocols are used, protocol parsing and encapsulation are not required

2. The client must do serialization and deserialization. This has nothing to do with communication

Copy the code

Caller sequence diagram


Time sequence diagram of the server


Caches manage multiple maps:

Which function calls are provided by the service provider by scanning jar package reflection to get the class and method names provided by the service

Copy the code

Protocol design

Data protocol

1, the Protocol Buffer

2. Segmentator, version number and Mesh message composition

Copy the code

1. There is a version number in the primary transmission protocol

For example, version 1 indicates RPC protocol

Version 2 indicates the mesh protocol

You can distinguish between the old protocol and the new protocol based on the version number

2. Multiple packets are separated by header and tail separators

3. The separator is 5 bytes

Copy the code

Mesh communication protocol

1. TCP long connection

Http1.1 or 2.0

Copy the code

Hybrid Cloud Deployment

1. Caller

A, sidecars + Service (Mesh)

B, the Service (RPC)

2. Service Party

A, sidecars + Service (Mesh)

B, the Service (RPC)

Copy the code

Access to the process


1. When the service is started, the Mesh service or common RPC service will register with the registry, and then the service type of the node will be known

2. The caller can pull down the service information to know the provider service type and then choose a different protocol to call

Copy the code

Small details

1. Fusing is done in mesh without the participation of the business side

2. The downstream retry times are the same. The service granularity is not the interface granularity

3. Health detection between proxies (Mesh) is distributed and the local routing table is updated once problems occur in the upstream or downstream of the proxy

4. Load balancing algorithm :Random, RR, Hash(consistency is mainly usedhashTo do)

(RR :(cyclic load)

The first request is routed to the first node,

The second request is routed to the second node,

The third request is routed to the third node,

The fourth request is routed to the first node

...).

Copy the code

Architecture in the future

2 Platform 1 Center 1 Trend



Service Mesh Decouples the platform and services

Container cloud elastic platform

Service governance platform (control center, registry, data collection center)

Artificial Intelligence (AI)

Copy the code

Call relationship – data collection storage method for service management platform

The service-caller perspective


Service provider: 5 million records per minute

Caller: 500,000

A total of 5.5 million

Copy the code

Storage solution

Plan 1


Scheme 2

Duplicate data is extracted as metadata


Plan 3

The actual call traffic is only 1/10 of scheme 1

Copy the code

,