Reading Notes: Distributed Systems

Let's review the two latest books I read: Designing Data-Intensive Application by Martin Kleppmann (DDIA for short) and Distributed Systems by Maarteen van Steen and Andrew S. Tanenbaum.

Both have a lot of similarities, they both make an academic description of the technologies, yet their content is very different. What do I mean by an academic description? The authors do not take position, or list the + and - of their subjects. They list the properties of their subjects: what happen when it fails? How many messages are exchanged? What kind of failures can we tolerate? etc. This approach is much more nuanced than the +/- approach. It gives deep knowledge in what make every technology unique. It provides the insights to make more informed decisions.


Do read Designing Data-Intensive Application, and enjoy the deep knowledge and design of systems.


This part is going to summarize the first 11 chapters of the book. It is going to be uselessly long, I advise you to skim through. If any part seems interesting to you, read the book, it's filled with many more details.

Chapter 1: Reliable, Scalable, and Maintainable Applications

Reliability: every system is subject to failures. Deal with it!

Scalability: how does Twitter scale the timeline. First solution: you can write every tweets in a table, and timelines are a joint of the tweet table with the follower table. But since writing and reading from the same table creates contention, you can design another approach. Each tweet is queued to be delivered to every follower.

                          [t3, t2, t1]<-------- follower 1
user --->[t4, t3, t2, t1]-[t4, t3, t2, t1]<---- follower 2
                          [t1, t2]<------------ follower 3

For instance, here where the user tweet, it is dispatched asynchronously to all its followers. The load, i.e the metric to quantify the performance of the system, is the speed at which a tweet is delivered to a follower. The problem of this approach is the write multiplication caused by the dispatch. The more follower you have, the more writes are dispatched to a lot of queues. If you have 1 million followers, 1 million writes are required to dispatch a single tweet.

Chapter 2: Data Models and Query Languages

There are multiple data models. Each have its own purpose and limitations.

The first model to exists, that almost disappear, is the hierarchical data. It was trumped by relationnal databases.

The core concepts of relational database are:

The core of a relational database is the query engine. It optimizes requests for you, and much effort has been put into it to make query fast.

The issue with the relational model is that it mismatches the object model of applications. Thus the need for complex ORM (Object-Relational Mapping).

The rise of document database allieviate this problem. Instead of normalizing data, you store JSON representation of your objects.

The queries for such model are simpler for read/write operations. They usually lack the expressiveness of SQL queries for more complex operations.

Finally the last data model is the graph-like data model.

You could use recursive RTE along with vertices/edges tables to simulate graph in SQL. The complexity of SQL queries make it hard to maintain such code.

This chapter covers a lot of details on graph languages, and shows their potency.

Chapter 3: Storage and Retrieval

This chapter solves the question: how to store data efficiently? This question is not easy to answer, let's see why.

First the quickest file operation is the append-only operation. Most database use this structure to store each version of an entry. Database use compaction to reduce the log. Say you have this log for instance:

mew: 1078
purr: 2103
purr: 2104
mew: 1079
mew: 1080
mew: 1081
purr: 2105
purr: 2106
purr: 2107
yawn: 511
purr: 2108
mew: 1082

The compaction process reduces it to the equivalent following:

yawn: 511
purr: 2108
mew: 1082

But then how do you retrieve data from the log? The naïve approach it to scan the file to find the latest entry. You cannot scale this approach.

Instead of scanning the file, you can maintain a hash index in memory containing the offset of each entry. This speeds up reads. You need to write to the log and to the index at the same time to guarantee consistency.

Hash tables have their limitations:

SSTables (Sorted String Tables) store keys in order in a sparse index. The corresponding entries are stored in blocks of x keys. For instance if x = 3:

Sparse Index in memory

hammer   -> 1049
handbag  -> 1130
handsome -> 1204
hangout  -> 1295
Sorted segment file

offset 1049: hammer:43, hammock: 123, hand: 321
offset 1130: handbag: 1928, handcuffs: 432, handful: 345
offset 1204: handsome: 24...

When looking for the entry for "handcuffs":

You can optimize this system to optimize reads. But how do you write in such structures? The solution is to use the append-only log as a buffer before writing in bulk. This is called LSMT (Log-Structured Merge-Tree) and widely used. Apache Cassandra, HBase, Google's BigTable, LevelDB engine in Riak all use this approach, as well as Elasticsearch and Lucene.

B-trees are also widely used in databases and are very suited for modern SSD hardware.

There are other type of index: multi-column index, storing entries in the index, etc. There are no clear winner and every index serve a purpose.

The author then separate two kind of workload: data analytics and transactional operations. Data analytics is a way of analysing the data to make sense of it all. Transactional workloads require highly available and low latency databases.

Relational databases use a row oriented storage: every entry is stored all together. On the other hand column oriented databases are very efficient for analytics workload. If you want to query the 90th percentile of the price of all products, in a row-oriented database you will have to scan every product. In a column-oriented database, every price is stored at the same location, making retrieval trivial.

This chapter is a gold mine of information to understand how database store and retrieve data.

Chapter 4: Encoding and Evolution

This chapter lists and compare major encoding format:

Also it covers the problem of evolution of schemas. To evolve a schema, you need:

Some operation on the schema are backward compatible, like adding a field. Removing a field on the other hand is not, so it needs to be handled carefully.

Format with schema offer an IDL (Interface Definition Language) that are useful for RPC (Remote Process Call). RPC has a lot of constraint, but have become necessary with service oriented architectures.

This chapters covers a lot of format and discussions around schema evolution and RPC are enlightening.

Chapter 5: Replication

Many reasons pushes data to be stored in a distributed manner:

The chapter covers replication issues.

There is synchronous replication: every write to the leader requires to be forwarded and acknowledged by the followers before being committed. Asynchronous replication is when only the leader confirms the write.

Asynchronous replication draws lesser latency, at the cost of weaker consistency guarantees. On the other hand synchronous replication suffers from weaker fault tolerence.

The chapter goes on explaining consistency levels you can expect from log replication. I already covered this in a previous post, so I won't talk about it here. Just know that the author is much better than I for explaining the different models.

Chapter 6: Partitionning

Usually partitionning is used along side with replication. Partitions have their own problems and they add up to replication problems.

The goal of partitions is to spread the data and the load across nodes. You can devise 2 main strategies to split data:

How to deal with secondary indexes when partitionning data. There are two approaches:

Another problem is skewed load and hot spots. Say your data is partitioned evenly. If only keys from the same partition are queried and modified, there is a hotspot in the partition. Choosing the right partition strategy is key to avoid this misbehavior.

Finally, when partitions are on different node, 3 strategies can be used to route requests.

The problem, known as service discovery, is solved differently by services. Kafka, HBase and SolrCloud use a routing tier, managed by ZooKeeper. Cassandra and Riak use a gossip protocol and let node manage the redirection to the correct node.

Chapter 7: Transactions

Transactions are tightly linked to consistency models. The chapter presents the strong (Serializibility) and the weaker models (Repeatable Read, etc.).

It also explains the different approach of Postgres for strong models. Postgres uses snapshot isolation for reads, and check that objects queried are not modified by another process when the transaction commits. If it is the case, it rollbacks the transaction, guaranteeing that there is no lost update. Stronger level of consistency are achieved by optimistically executing concurrent requests.

On the other hand MySQL uses a stronger 2 phase locking to make sure no two process access data.

Finally there is FoundationDB approach: make sure requests are treated by a single thread at any point. By partitionning data to every thread, this approach is easy to implement and quite efficient.

Each paradigm has pros and cons, and it's up to your application to use the right solution.

This chapter covers a lot more than that, but I don't have enough place to write about every goodness in this book.

Chapter 8: The Trouble with Distributed Systems

Distributed systems are subjects to faults, failure and all sort of bad things.

Network can fail.
Network can be slow.
Network can have unbounded latency.

Clocks in computers are based on quartz, which drift from other quartz. Google estimate that clocks drift up to 17 seconds per day between quartz.

NTP is a protocol to correct the precision of clocks by comparing local time to a GPS clock or atomic clock. Even if NTP is useful, it can't be used to order the operation of distributed systems. Garbage collection in language environment or Virtual machine pauses make clocks even less reliable.

Truth is difficult to achieve in distributed systems as well. Imagine: a lock is held by a process which craches. Another process claims the lock but receives no answer, what should it do?

Even if a majority of process agree, Byzantine faults can make the agreement fail. Although not common they are things to think about.

Chapter 9: Consistency and Consensus

Major problems in distributed systems: agreement, atomic broadcast, totally ordered dispatch, uniqueness, membership detectors can be reduced to a single problem: consensus. This is why consensus is so important to this field of computing.

If a consensus is reach by processes in a distributed system, the system can act like it would be a single process. This is the graal of every distributed systems. Unfortunately consensus is very difficult to achieve by faulty processes and unreliable network.

The first and most influential solution to this problem is the Paxos protocol.

The chapter also explains 2 phase commit and how failing coordinator is hard to recover from. XA protocol is not implemented by a lot of systems, therefore it's not really a killer feature.

Consensus solution are very difficult and expensive to achieve, this is why developers should choose a weaker model, or rely on battle-proofed solutions: ZooKeeper.

Chapter 10: Batch Processing

This chapter explains the approach of MapReduce, and how it sets the foundation of big data processing.

Like the unix pipe operator, MapReduce approach is to offer a coherent way of pipelining operation on data. HDFS the Hadoop file system stores raw dataset as files. Each step on the pipeline transforms the data and save its result to HDFS.

It's easy to replay tasks, since every step is saved on disks. Also since data is immutable, you can perform many steps in parallel on the same dataset.

The chapter goes on to describe how MapReduce tasks are performed. It also describes strategies to join large datasets.

The problem of storing all information of intermediary steps is not efficient when many steps are needed. To solve this issue, well known solutions like Spark, Flink or Tez exist. They are known as dataflow engines.

They need special care. Since intermediary steps are not written to disk, a failure in the pipeline can have dramatic effect.

This chapter is really enlightening if you've never met the large data warehouse of some companies. When dealing with petabytes of data, you cannot use relationnal model and MapReduce solves the problems quite elegantly.

Chapter 11: Stream Processing

This last chapter is about stream processing: using Kafka as a messaging system. Kafka is a hybrid approach that can do the same workload has MapReduce thanks to storing event in log and offering replay.

This chapter covers how to use Kafka to do CDC (Change Data Capture). Many solutions exist that I'm not gonna cover.

The author mentions here that CDC is the core concept of event sourcing, but on the application level.

The word that I learn in this chapter is derived data. The source of truth can be an relationnal database, MapReduce, CDC or event sourcing allow to generate derived data. The derived data can be an index, an analytics dashboard, etc. CDC is a very powerful tool to solve problems like:

But then can also be used to maintain materialized view, either for MapReduce or application usage.

To make stream processing reliable and resilient to faults, some solutions are possible. You can use checkpoints or use smaller batch that can be retried. You can also have idempotent jobs that can safely process duplicated messages.

This summary does not capture all details of the book. It is annotated by a lot of papers. You can dig sujects so deep that you'll find yourself in the heart of the earth.

Distributed systems

Contrary to the last book, I won't cover every chapter because it did not capture all my attention. It goes in much more details. For instance there are a detailed description of the Paxos protocol. It explains in details the internals of CDN (Content Delivery Network).

The takeaways for me are:

The book is available for free only, but i recommend you to read chapters 6-7-8 in priority. Because authors goes in much detail quickly, I felt lost when it concerned network disposition. Paradoxically, it tries to use simple description, but rarely mentions industry usage of the solutions. It makes it difficult to picture what problems we try to solve.


Both books have an academic approach: they list a number of paradigm and try to compare them faithfully. Both achieve it with brillance although I would mostly recommend DDIA if you are an application developer.

Happy new year everyone!