Distributed systems are composed of many smaller systems. In this section, we explore three fundamental composition patterns in detail:
- Load balancer with multiple backend replicas
- Server with multiple backends
- Server tree
1.3.1 Load Balancer with Multiple Backend Replicas
The first composition pattern is the load balancer with multiple backend replicas. As depicted in Figure 1.1, requests are sent to the load balancer server. For each request, it selects one backend and forwards the request there. The response comes back to the load balancer server, which in turn relays it to the original requester.
Figure 1.1: A load balancer with many replicas
The backends are called replicas because they are all clones or replications of each other. A request sent to any replica should produce the same response.
The load balancer must always know which backends are alive and ready to accept requests. Load balancers send health check queries dozens of times each second and stop sending traffic to that backend if the health check fails. A health check is a simple query that should execute quickly and return whether the system should receive traffic.
Picking which backend to send a query to can be simple or complex. A simple method would be to alternate among the backends in a loop—a practice called round-robin. Some backends may be more powerful than others, however, and may be selected more often using a proportional round-robin scheme. More complex solutions include the least loaded scheme. In this approach, a load balancer tracks how loaded each backend is and always selects the least loaded one.
Selecting the least loaded backend sounds reasonable but a naive implementation can be a disaster. A backend may not show signs of being overloaded until long after it has actually become overloaded. This problem arises because it can be difficult to accurately measure how loaded a system is. If the load is a measurement of the number of connections recently sent to the server, this definition is blind to the fact that some connections may be long lasting while others may be quick. If the measurement is based on CPU utilization, this definition is blind to input/output (I/O) overload. Often a trailing average of the last 5 minutes of load is used. Trailing averages have a problem in that, as an average, they reflect the past, not the present. As a consequence, a sharp, sudden increase in load will not be reflected in the average for a while.
Imagine a load balancer with 10 backends. Each one is running at 80 percent load. A new backend is added. Because it is new, it has no load and, therefore, is the least loaded backend. A naive least loaded algorithm would send all traffic to this new backend; no traffic would be sent to the other 10 backends. All too quickly, the new backend would become absolutely swamped. There is no way a single backend could process the traffic previously handled by 10 backends. The use of trailing averages would mean the older backends would continue reporting artificially high loads for a few minutes while the new backend would be reporting an artificially low load.
With this scheme, the load balancer will believe that the new machine is less loaded than all the other machines for quite some time. In such a situation the machine may become so overloaded that it would crash and reboot, or a system administrator trying to rectify the situation might reboot it. When it returns to service, the cycle would start over again.
Such situations make the round-robin approach look pretty good. A less naive least loaded implementation would have some kind of control in place that would never send more than a certain number of requests to the same machine in a row. This is called a slow start algorithm.
1.3.2 Server with Multiple Backends
The next composition pattern is a server with multiple backends. The server receives a request, sends queries to many backend servers, and composes the final reply by combining those answers. This approach is typically used when the original query can easily be deconstructed into a number of independent queries that can be combined to form the final answer.
Figure 1.2a illustrates how a simple search engine processes a query with the help of multiple backends. The frontend receives the request. It relays the query to many backend servers. The spell checker replies with information so the search engine may suggest alternate spellings. The web and image search backends reply with a list of web sites and images related to the query. The advertisement server replies with advertisements relevant to the query. Once the replies are received, the frontend uses this information to construct the HTML that makes up the search results page for the user, which is then sent as the reply.
Figure 1.2: This service is composed of a server and many backends.
Figure 1.2b illustrates the same architecture with replicated, load-balanced, backends. The same principle applies but the system is able to scale and survive failures better.
This kind of composition has many advantages. The backends do their work in parallel. The reply does not have to wait for one backend process to complete before the next begins. The system is loosely coupled. One backend can fail and the page can still be constructed by filling in some default information or by leaving that area blank.
This pattern also permits some rather sophisticated latency management. Suppose this system is expected to return a result in 200 ms or less. If one of the backends is slow for some reason, the frontend doesn’t have to wait for it. If it takes 10 ms to compose and send the resulting HTML, at 190 ms the frontend can give up on the slow backends and generate the page with the information it has. The ability to manage a latency time budget like that can be very powerful. For example, if the advertisement system is slow, search results can be displayed without any ads.
To be clear, the terms “frontend” and “backend” are a matter of perspective. The frontend sends requests to backends, which reply with a result. A server can be both a frontend and a backend. In the previous example, the server is the backend to the web browser but a frontend to the spell check server.
There are many variations on this pattern. Each backend can be replicated for increased capacity or resiliency. Caching may be done at various levels.
The term fan out refers to the fact that one query results in many new queries, one to each backend. The queries “fan out” to the individual backends and the replies fan in as they are set up to the frontend and combined into the final result.
Any fan in situation is at risk of having congestion problems. Often small queries may result in large responses. Therefore a small amount of bandwidth is used to fan out but there may not be enough bandwidth to sustain the fan in. This may result in congested network links and overloaded servers. It is easy to engineer the system to have the right amount of network and server capacity if the sizes of the queries and replies are consistent, or if there is an occasional large reply. The difficult situation is engineering the system when there are sudden, unpredictable bursts of large replies. Some network equipment is engineered specifically to deal with this situation by dynamically provisioning more buffer space to such bursts. Likewise, the backends can rate-limit themselves to avoid creating the situation in the first place. Lastly, the frontends can manage the congestion themselves by controlling the new queries they send out, by notifying the backends to slow down, or by implementing emergency measures to handle the flood better. The last option is discussed in Chapter 5.
1.3.3 Server Tree
The other fundamental composition pattern is the server tree. As Figure 1.3 illustrates, in this scheme a number of servers work cooperatively with one as the root of the tree, parent servers below it, and leaf servers at the bottom of the tree. (In computer science, trees are drawn upside-down.) Typically this pattern is used to access a large dataset or corpus. The corpus is larger than any one machine can hold; thus each leaf stores one fraction or shard of the whole.
Figure 1.3: A server tree
To query the entire dataset, the root receives the original query and forwards it to the parents. The parents forward the query to the leaf servers, which search their parts of the corpus. Each leaf sends its findings to the parents, which sort and filter the results before forwarding them up to the root. The root then takes the response from all the parents, combines the results, and replies with the full answer.
Imagine you wanted to find out how many times George Washington was mentioned in an encyclopedia. You could read each volume in sequence and arrive at the answer. Alternatively, you could give each volume to a different person and have the various individuals search their volumes in parallel. The latter approach would complete the task much faster.
The primary benefit of this pattern is that it permits parallel searching of a large corpus. Not only are the leaves searching their share of the corpus in parallel, but the sorting and ranking performed by the parents are also done in parallel.
For example, imagine a corpus of the text extracted from every book in the U.S. Library of Congress. This cannot fit in one computer, so instead the information is spread over hundreds or thousands of leaf machines. In addition to the leaf machines are the parents and the root. A search query would go to a root server, which in turn relays the query to all parents. Each parent repeats the query to all leaf nodes below it. Once the leaves have replied, the parent ranks and sorts the results by relevancy.
For example, a leaf may reply that all the words of the query exist in the same paragraph in one book, but for another book only some of the words exist (less relevant), or they exist but not in the same paragraph or page (even less relevant). If the query is for the best 50 answers, the parent can send the top 50 results to the root and drop the rest. The root then receives results from each parent and selects the best 50 of those to construct the reply.
This scheme also permits developers to work within a latency budget. If fast answers are more important than perfect answers, parents and roots do not have to wait for slow replies if the latency deadline is near.
Many variations of this pattern are possible. Redundant servers may exist with a load-balancing scheme to divide the work among them and route around failed servers. Expanding the number of leaf servers can give each leaf a smaller portion of the corpus to search, or each shard of corpus can be placed on multiple leaf servers to improve availability. Expanding the number of parents at each level increases the capacity to sort and rank results. There may be additional levels of parent servers, making the tree taller. The additional levels permit a wider fanout, which is important for an extremely large corpus. The parents may provide a caching function to relieve pressure on the leaf servers; in this case more levels of parents may improve cache effectiveness. These techniques can also help mitigate congestion problems related to fan-in, as discussed in the previous section.