One of the big changes in RavenDB is the new search engine, Corax. We want to replace Lucene with a purpose built search engine, capable of doing everything that we can do with Lucene, but far faster.

In this post, I want to discuss a particular detail of the implementation. Managing the posting list. In information retrieval, the idea of a posting list is that we have a list of document ids that match a particular term. I’m ignoring the details of how we create or manage the list. The basic idea is that we have a need to store a potentially large number of document ids, update them on the fly as well query them. Lucene manages that by creating immutable segments, but that design decision doesn’t match the way we want to do things in Corax. The problem with immutable segments is that you’ll eventually need to run compaction, which can be really expensive.

As a database, we already have a really good data structure that matches most of those requirements, it turns out. A B+Tree can do a pretty good approximation of a posting list, but it does have a problem. It’s not efficient in terms of space usage. A document id is a 64 bits integer, and we can make a number of assumptions about it. Therefore, I create a dedicated B+Tree like structure to hold the posting list. This B+Tree is called a Set inside of Voron, and it can hold any number of uint64 values. Internally, this is managed as two separate types of pages. The Branch pages (which are fairly standard B+Tree branches) and the Leaf pages.

Here is the first version of the Set Leaf Page implementation:

Let’s figure out what is going on here. The page has a notion of a base. That means all the documents IDs that have the same upper 33 bits. Basically, inside a single page, all the IDs are in the same 2 billion range. That means that even though the document ids are 64 bits, in the context of a single page, we can treat them as 32 bits integers. That turns out to be important, since most integer compression routines work on 32 bits integers.

How does that work? We have a section in the page that is dedicated to raw values, and we insert values into that section until it is full. Whenever that happens, we compress the raw values section using PForDelta compression. The page will then contain multiple compressed segments and a single raw values segment. Each compressed segment is non-overlapping with the other compressed segments (but may overlap with the raw values). PForDelta is really good in compressing integers, especially if it is able to figure out patterns in your data. And documents IDs in Corax are explicitly designed to have common patterns so it will be able to take advantage of this behavior. When we read the entries from the page, we merge the compressed segments with the raw values and return the full details.

The code isn’t particularly simple, but has a fairly straightforward approach to the problem once you understand the design.

One thing that I haven’t touched is the notion of removals. That is an important concept, and we handle that in an interesting way. Remember that I said that the baseline for the page is the upper 33 bits? That is because the numbers inside the page are 31 bits in length, we reserve the top bit to mark a value as a tombstone marker.  In other words, when we write to the raw values, we can have either a new value or a removed value, distinguished using the uppermost bit.

When we fill the raw values segment, we compress it alongside the relevant compressed segments. At that point, we’ll filter out the removed values. This is very similar to the segments that Lucene is using, but we do that on a page boundary (8KB), not across all values.

We are able to push a lot of values into a single page. We see typically thousands to tens of thousands of documents IDs fitting into a single 8KB page. That is pretty amazing, since even if you have a posting list that has millions of entries, the actual disk reads are minimal.

The design was with us throughout most of the development of Corax, but we ran into a serious issue with the way it works when we started to build the query optimizer for Corax.

That is an interesting statement, I think you’ll agree. What is the relation between a (very) low-level design of the on-disk data format and the result of the query optimizer?

Well, it turns out that one of the things that we need to know for the query optimizer is: “How big is this posting list?”

This question is actually really important to be able to generate optimal queries. And given the structure we have, we can provide a good answer to that, most of the time, but not always.

Why is that?

The problem is what happens when we want to remove a value from the set, or add an existing value. If the value already exists inside a compressed segment, we don’t open the compressed segement (which will require re-writing it from scratch), so we record an addition that is actually spurious. Conversely, if we try to remove a value that isn’t in the set, we’ll wrongly decrement the number of entries in the posting list, leading to issues with a mismatch between the record number of entries and the actual number we have in the posting list.

That was very confusing to figure out, I’ll admit. It was also really hard to fix, but I’ll address that in the next post in the series.

I’ve been calling myself a professional software developer for just over 20 years at this point. In the past few years, I have gotten into teaching university courses in the Computer Science curriculum. I have recently had the experience of supporting a non-techie as they went through a(n intense) coding bootcamp (aiming at full stack / front end roles). I’m also building a distributed database engine and all the associated software.

I list all of those details because I want to make an observation about the distinction between fundamental and transient knowledge.

My first thought is that there is so much to learn. Comparing the structure of C# today to what it was when I learned it (pre-beta days, IIRC), it is a very different language. I had literally decades to adjust to some of those changes, but someone that is just getting started needs to grasp everything all at once. When I learned JavaScript you still had browsers in the market that didn’t recognize it, so you had to do the “//<!—” trick to get things to work (don’t ask!).

This goes far beyond mere syntax and familiarity with language constructs. The overall environment is also critically important. One of the basic tasks that I give in class is something similar to: “Write a network service that would serve as a remote dictionary for key/value operations”.  Most students have a hard time grasping details such as IP vs. host, TCP ports, how to read from the network, error handling, etc. Adding a relatively simple requirement (make it secure from eavesdroppers) will take it entirely out of their capabilities.

Even taking a “simple” problem, such as building a CRUD website is fraught with many important details that aren’t really visible. Responsive design, mobile friendly, state management and user experience, to name a few. Add requirements such as accessibility and you are setting the bar too high to reach.

I intentionally choose the examples of accessibility and security, because those are “invisible” requirements. It is easy to miss them if you don’t know that they should be there.

My first website was a PHP page that I pushed to the server using FTP and updated live in “production”. I was exposed to all the details about DNS and IPs, understood exactly that the server side was just a machine in a closet, and had very low levels of abstractions. (Naturally, the solution had no security or any other –ities). However, that knowledge from those early experiments has served me very well for decades. Same for details such as how TCP works or the basics of operating system design.

Good familiarity with the basic data structures (heap, stack, tree, list, set, map, queue) paid itself many times over. The amount of time that I spent learning WinForms… still usable and widely applicable even in other platforms and environments. WPF or jQuery? Not so much.

Learning patterns paid many dividends and was applicable on a wide range of applications and topics.

I looked into the topics that are being taught (both for bootcamps and universities) and I understand why in many cases, those are being skipped. You can actually be a front end developer without understanding much (if at all) about networks. And the breadth of details you need to know is immense.

My own tendency is to look at the low level stuff, and given that I work on a database engine, that is obviously quite useful. What I have found, however, is that whenever I dug deep into a topic, I found ways to utilize that knowledge at a later point in time. Sometimes I was able to solve a problem in a way that would be utterly inconceivable to me previously. I’m not just talking about being able to immediately apply new knowledge to a problem. If that were the case, I would attribute that to wanting to use the new thing I just learned.

However, I’m talking about scenarios where months or years later I ran into a problem, and was then able to find the right solution given what was then totally useless knowledge.

In short, I understand that chasing the 0.23-alpha-stage-2.3.1-dev updates on the left-pad package is important, but I found that spending time deep in the stack has a great cumulative effect.

Joel Spolsky wrote about leaky abstractions, that was 20 years ago. I remember reading that blog post and grokking that. And it is true, being able to dig one or two layers down from where you usually live has a huge amount of leverage on your capabilities.

A user came to us with an interesting scenario. They have a RavenDB cluster, which is running in a distributed manner. At some point, we have a user that creates a document inside of RavenDB as well as posts a message using SQS (Amazon queuing system) to be picked up by a separate process.

The flow of the system is shown below:

The problem they run into is that there is an inherent race condition in the way they work. The backend worker that picks up the messages may use a different node to read the data than the one that it was written to.

RavenDB uses asynchronous replication model, which means that if the queue and the backend workers are fast enough, they may try to load the relevant document from the server before it was replicated to it. Amusingly enough, that typically happens on light load (not a mistake, mind). On high load, the message processing time usually is sufficient to delay things for replication to happen. In light load, the message is picked up immediately, exposing the race condition.

The question is, how do you deal with this scenario? If this was just a missing document, that was one thing, but we also need to handle another scenario. While the message is waiting to be processed in the queue, it may be updated by the user.

So the question now is, how do we handle distributed concurrency in a good manner using RavenDB.

The answer to this question is the usage of change vectors.  A change vector is a string that represents the version of a document in a distributed environment. The change vector is used by RavenDB to manage optimistic concurrency.

This is typically used to detect changes in a document behind our backs, but we can use that in this scenario as well. The idea is that when we put the message on the queue, we’ll include the change vector that we got from persisting the document. That way, when the backend worker picks up the message and starts using it, the worker can compare the change vectors.

If the document doesn’t exist, we can assume that there is a replication delay and try later. If the document exists and the change vector is different, we know the document was modified, and we may need to use different logic to handle the message in question.

A database indexing strategy is a core part of achieving good performance. About 99.9% of all developers have a story where adding an index to a particular query cut the runtime from seconds or minutes to milliseconds. That percentage is 100% for DBAs, but the query was cut from hours or days to milliseconds.

The appropriate indexing strategy is often a fairly complex balancing act between multiple competing needs. More indexing means more I/O and cost on writes, but faster reads. RavenDB has a query optimizer engine that will analyze your queries and generate the appropriate set of indexes on the fly, without you needing to think much about it. That means that RavenDB will continuously respond to your operational environment and changes in it. The end result is an optimal indexing strategy at all times.

This automatic behavior applies only to automatic indexes, however. RavenDB also allows you to define your own indexes and many customers run critical business logic in those indexes. RavenDB now has a feature that aims to help you manage/organize your indexes by detecting redundant definitions & unqueried indexes which can be removed or merged.

The index cleanup feature is now exposed in the Studio (since build 5.4.5):

When you select it, the Studio will show you the following options:

You can see that RavenDB detected that two indexes can be merged into a single one, and additionally there are some indexes that haven’t been used in a while or have been completely superseded by other indexes.

RavenDB will even go ahead and suggest the merged index for you:

The idea is to leverage RavenDB’s smarts so you won’t have to spend too much time thinking about index optimization and can focus on the real value-added portions of your system.

I recently talked about the beta release of the RavenDB PHP client a few months ago. I’m now really pleased to announce that we have just released the official RavenDB PHP client.

Here is what this looks like:

I had a lot of fun in this webinar, showing off some of RavenDB’s capabilities in really complex distributed systems:

We have just released a new stable release of the RavenDB Python client API. This puts the Python client API for RavenDB on the same level as our other clients, including support for subscriptions, cluster wide transactions, compare exchange, conditional loading, and much more.

We also improved the ergonomics of the API and integration with the IDE.

Here is an example of writing a non-trivial query using the API, tell us what you think and what you are doing with RavenDB & Python.