|
|
|
|
|
by nonbirithm
2311 days ago
|
|
Here's an explanation from CloudFlare as to the root cause. [0] > I am the engineering manager for the DDoS protection team and this morning at 11:06 UTC we tweaked a rule that affected one of our signals. The signal relates to the HTTP referer header, and we have a piece of code that looks at invalid referer headers. In this case we tweaked it to include not just "obvious garbage" but "anything that does not conform to the HTTP specification"... i.e. is the referer a URI? If not then it contributes to knowledge about bad traffic. > So... why did this impact npmjs.org? It turns out that a lot of NPM traffic sends the referer as "install" which is invalid according to the HTTP specification. As NPM is also a heavily trafficked site this resulted in the DDoS systems picking this up and treating the traffic as a HTTP flood and determining that a rate-limit should be applied. > When we noticed that NPM was seeing an increase in HTTP 429s (as seen on Twitter) we contacted NPM and started an internal investigation. As soon as we identified the root cause we reverted the change, which was at 13:00 UTC. > We'll note that NPM and 1 other site use the referer for purposes outside the HTTP spec and we'll update our systems to ensure that this does not happen again. Additionally we'll improve our monitoring around changes of this nature so that we can discover impact sooner and roll back automatically. [0] https://github.com/npm/cli/issues/836#issuecomment-587019096 |
|