Recently, I've been helping a few video stations to do architecture upgrades, and I've stepped all over the CDN pits. Remember last year, our self-built video service was DDoS hit collapse, the boss stared at the downtime statement face is green - 48 consecutive hours of inaccessibility, user turnover rate directly soared to 35%. later I realized that these days to do the video station, there is no reliable high-defense CDN is simply running naked on the battlefield.
Video traffic is simply not on the same order of magnitude as regular web pages. A 1080p video bandwidth consumption may be hundreds of times the graphic page, not to mention the sudden traffic and malicious attacks. I found that the actual test, ordinary CDN encountered more than 100G DDoS basically on the spot on the street, and those who claimed that “unlimited protection” business, really encountered an attack on the first to give you the first to limit the flow of them.
Don't believe those theoretical indicators on paper.The “T-level protection” and “intelligent scheduling” are all nonsense without being tested by real traffic. What "T-level protection" "intelligent scheduling", not through the real traffic test are rogues. Last year, when testing a vendor, their console shows that everything is normal, but the actual users complained about the lag. Later, ffmpeg pull stream test only to find that their node scheduling algorithms can not carry cross-province fluctuations.
First of all, the video CDN must solve the three core problems: bandwidth costs, DDoS protection, lag optimization. Bandwidth cost, some vendors play the “overselling game” - the promise of low-priced bandwidth, but really run up to give you a speed limit. DDoS protection is even worse, some small factories are hit on the Null route to your IP, which is equal to directly to you to pull out the network cable.
After testing three major service providers, here's the dry goods directly:
CDN5's resilient protection does topThe client's website was hit by 800Gbps last week. Last week, the customer site was hit to 800Gbps, automatically triggered their Anycast cleaning system, and ultimately landed to the source of the traffic less than 5G. I deliberately look at their server room topology, more than 60 cleaning centers around the world linkage, the Hong Kong node is also able to resist LDAP reflection attacks. Although the price is more expensive than ordinary CDN 30%, but compared to the downtime loss is simply a hair.
Configuration example (nginx linkage scheme):
I'm impressed with CDN07's smart routingThe video dedicated line supports BGP dynamic routing. Their video dedicated line supports BGP dynamic routing, which simply means that it can bypass congested backbone nodes. The measured latency from Guangzhou to Los Angeles is stabilized at about 140ms, and the packet loss rate is controlled at below 0.5%. The most practical is their warm-up function - push popular videos to the edge nodes in advance, and respond directly locally to sudden requests.
Here's a sample preheat API call I'm using:
08Host's cost-effective route for startups. Although the protection is slightly weaker (up to 300Gbps cleaning), but their price is only half of the big players. Particularly to praise is their console experience, bandwidth fluctuation charts can be accurate to every second, but also comes with CC attack recognition rule base. I gave them a work order suggestion to increase the TCP acceleration function, I did not expect to update it two weeks later.
Comparison of core data from the three vendors:
Key Recommendation:Don't just look at the price to choose a program. Once I chose a small vendor for the cheapest price, and the result was that during the Christmas promotion period, the traffic skyrocketed, and their nodes were directly overloaded and down. Later monitoring found that some vendors of the “unlimited traffic” in fact hidden peak limit - more than the agreed bandwidth will directly lose packets.
Now our architecture using dual CDN hot standby program: the main line with CDN5 to do protection and acceleration, standby line with 08Host to carry burst traffic. Once a month, we do a real failure drill to simulate a single point of failure switching process. This program allows us to improve our SLA from 99.5% to 99.99%, although the cost increased by 40%, but the user retention rate increased by more than two times.
Finally said a tearful lesson: once superstitious international CDN, the results of their China node landing quality is extremely poor. During the evening peak, the delay of southern users directly soared to 800ms+. Only later did I understand that the selection of CDN must look at the localization ability - business license, ICP record, local technical support team, the three missing one may step on the pit.
If you really want to recommend, large video stations choose CDN5 + self-built node hybrid architecture, small and medium-sized enterprises with CDN07 completely sufficient, startup teams start from 08Host first. Remember, there is no universal antidote, only the program suitable for the current stage. Every time I see those brainless push a CDN soft I want to laugh - out of the business scene to talk about technology selection is blind bullshit.
(Test data as of March 2024, the specific policy is subject to the manufacturer's latest announcement. (If you want to know how to press the price, you can private me, the water in this business is very deep, at least you can negotiate a discount of 15%-20%)
