Two important traits of building reliable distributed systems

Designing a distributed system is hard enough. Even harder to design a distributed system that is reliable. There are many best practices that you can follow to make a reliable distributed system. Based on an issue that I recently troublehooted, there are a couple of them that I think are critical:

  • Enabling TCP keep-alive between the processes if you are using TCP
  • Performing all IO operations with a time out
My advise is based on my experience in Linux. But I think it should be applicable to other operating systems as well. When a host goes down with a kernel panic, none of the established connections are closed by sending a FIN or RESET packet. This cauases trouble that the peer process doesn't know about the other end of the communication being gone. When you enable TCP keep-alive, the kernel sends a zero length packet as per the configuration. Hence if the peer has died due to kernel panic, the zero length packet will not be ACKed. Thus, the peer death is detected. If the peer has died and rebooted, when the zero length packet is sent, the peer responds with RESET packet. Thus the best way to detect peer death is enabling keep-alive.

For Linux, to enable keep-alive, please refer to this page. The most important thing is after you open the socket, you should use setsockopt() to set the SO_KEEPALIVE option.

Using timeouts is one way to detect that the service quality is degrading. For e.g. when you try to fetch data from an external system, the external system might be choking due to too many requests and may not be as responsive as it should be. But if you build retry mechanisms based on timeouts, you have to be extra careful. Because sending more requests to an external system might make the situation worse than it should be.

Comments

Post a Comment

Popular posts from this blog

Gotchas with DBCP

I had to make use of DB connection pool in one of the applications that I was implementing. Since I am new to Java, everyone in the team was suggesting why don't you make use of Commons DBCP. Being a fan of reuse, I decided to make use of DBCP and started reading about the parameters that are exposed so that I can customize DBCP to fit my needs. While reading through the available configuration parameters , I was thinking, "I don't need to check the connection to be bad on every borrow or return. It is sufficient if I check for being bad only when the connection is sitting idle in the pool." So I set the testWhileIdle to be true and set the validationQuery to be "SELECT 1 FROM DUAL". I was thinking that the configuration that I had was the optimal for the situation in hand. But one thing I overlooked was, the default value for the testOnBorrow is true. So the moment you set the validationQuery to be any non-null value, every connection you borrow from the po
Read more

A note on Java's Calendar set() method

Remember that the Calendar's internal fields include year, month, date, hour, minutes, seconds, milliseconds and time zone. Whenever you are calling a set() method with multiple fields, like set(year, month, date), it will not affect the rest of the fields. Remember that there is no set() method with multiple fields available to set the milliseconds. If you would like to set the milliseconds, you must use set(Calendar.MILLISECOND, value). Likewise, if you are planning to set all the fields, its a good idea to reset all the fields using clear() method. This will clear milliseconds as well. Most of the times, millisecond field may not be of interest to you. But if you are going to use the UTC milliseconds, by calling getTimeInMillis(), then make sure you set the right values for milliseconds as well.
Read more

The mysterious ORA-03111 error

Recently one of the applications that I developed started throwing exceptions, that had the following message: SQL state [72000]; error code [1013]; ORA-03111: break received on communication channel When I googled around, I couldn't come across anything useful. Sadly enough most of the sites just showed the documentation for that error, without any explanation from anyone experiencing that issues. So here you go, with the best possible explanation that I could come up with. My application sets two things on the connection that is throwing this exception: It sets the fetchSize to be 2500 rows It sets the query timeout to be 10 seconds The database server and the application are separated over a long latency network (actually there is a NetEm box that emulates the long latency between these two boxes) which has a latency characteristic of 50+/-5 milliseconds. This is the whole setup. It is important to understand how the timeout is handled by the Oracle client (in my case JDBC clien
Read more