Let's talk about eBay's best scalability practices.

Many people think that the technical lectures on eBay in qcon (Beijing) are good, but for me, the impact is not that great. This was eBay one or two years ago. "> PPT. However, it is much better than Amazon's Jeff Barr. If you want to hold a meeting in the future, please invite Jeff Barr to discuss the sales document. I guess I am sorry.

However, the PPT of eBay is updated a bit.

1) Partition (everything)

It is a partition. It cannot be equivalent to a partition in Oracle. It is good to understand the sharding. For more information, see the sharing nothing technology, which I wrote earlier ). It should be emphasized that slice is suitable for processing data only when the data size is indeed large. If a single node is sufficient to cope with it, Do not impersonate it.

In terms of the sharding mode, eBay uses function segmentation (functional segmentation) and horizontal segmentation (Server Load balancer) to infer that all sessions are stateless.

2) asynchrony everywhere)

In fact, the excessive pursuit of "synchronization" design for any website is still relatively bad. It is designed from the perspective of data that users can observe, and can be done asynchronously to the maximum extent.

There are two examples of eBay: event queue and message multicast ). The former is basically a producer-consumer model. The latter is mainly used in the search architecture.

Taking note of the message bus shown in the figure, This is the artery in eBay's entire architecture. It is estimated that technical details will not be easily reviewed.

3) Automation)

The automation here provides two examples: one for O & M and the other for machine learning. This is the strength of the speaker Randy Shoup.

EBay's automation, in another article a year ago, gave a glimpse of something. This article was not taken seriously by more people at the beginning. For details, see eclipse at eBay. We can see that eBay can do so well in automation (at least dare to say it) is not an overnight task.

4) fault detection and backtracking (remember everything fails)

Better failure detection mechanism: monitors logs that exceed 2 TB per day and generates judgments or warnings based on relevant events in the logs. This seems simple, but it still requires a little skill and strategy to implement it. What's important is that we need to constantly improve it based on the feedback of the results.

Perfect rollback: Any service is identified by the tag in the service configuration and does not require rollback. (I personally think this is very difficult, especially when upgrading)

Graceful degradation: it is easy to mark the application as "marks down (offline )"

5) embrace inconsistency)

With the CAP principle, Cheng Li described his image as a hat trick, which is very accurate. Speaking of consistency, since amazon "> CTOAs soon as Werner vogels eventually consistent came out, basically I don't need to talk nonsense. This is the story of transaction processing. Let's take a look at it at home.

Ebay also has its own absolute principle: there is absolutely no distributed transaction (two-phase commit), through the state machine and operation order to minimize inconsistency, through asynchronous events (Message bus ?) Achieve final consistency.

-- EOF --

Other News: Amazon "> CTOWerner vogels may attend the Grand Prix in Hangzhou in March.

From: http://www.dbanotes.net/arch/best_practices_for_scaling_websites_lessons_from_ebay.html