Jay Kreps - I [heart symbol] logs: event data, stream processing, and data integration

When most people think about logs they probably think about something that looks like .

Every programmer is familiar with this kind of loga series of loosely structured requests, errors, or other messages in a sequence of rotating text files.

This type of log is a degenerative form of the log concept I am going to describe. The biggest difference is that this type of application log is mostly meant for humans to read, whereas the logs Ill be describing are also for programmatic access.

Actually, if you think about it, the idea of humans reading through logs on individual machines is something of an anachronism. This approach quickly becomes unmanageable when many services and servers are involved. The purpose of logs quickly becomes an input to queries and graphs in order to understand behavior across many machines, something that English text in files is not nearly as appropriate for as the kind of structured log Ill be talking about.

Figure 1-1. An excerpt from an Apache log

The log Ill be discussing is a little more general and closer to what in the database or systems world might be called a commit log or journal. It is an append-only sequence of records ordered by time, as in .

Figure 1-2. A structured log (records are numbered beginning with 0 based on the order in which they are written)

Each rectangle represents a record that was appended to the log. Records are stored in the order they were appended. Reads proceed from left to right. Each entry appended to the log is assigned a unique, sequential log entry number that acts as its unique key. The contents and format of the records arent important for the purposes of this discussion. To be concrete, we can just imagine each record to be a JSON blob, but of course any data format will do.

The ordering of records defines a notion of time since entries to the left are defined to be older then entries to the right. The log entry number can be thought of as the timestamp of the entry. Describing this ordering as a notion of time seems a bit odd at first, but it has the convenient property of being decoupled from any particular physical clock. This property will turn out to be essential as we get to distributed systems.