Category Archives: Data

Downstream Wikipedia link usage and migration to Wikidata

MusicBrainz has linked to Wikipedia for many years and we now have links to Wikidata as well. Wikidata, however, acts as a central repository for Wikipedia links, so it does not make sense for MusicBrainz to maintain its own separate set of Wikipedia links, especially since Wikipedia URLs are not very stable (because of page moves and deletions) and require a lot of maintenance. Most of our data with Wikipedia links is now also linked to Wikidata, so we plan to start removing Wikipedia links where we have a Wikidata link which has the same Wikipedia link.

What this means for downstream data users:

If you use Wikipedia links, we will provide Wikidata links but you will need to fetch the Wikipedia links you want from Wikidata separately. Wikidata has information on ways to access their data at

We plan to start removing the links after the schema change this month, starting with the less common languages and entity types. It will take a while to work through the existing links, so we don’t expect to start removing English links from artists until after the Autumn schema change.

We recognise that some people may have code which depends on these links – if you’re using these links and the above sounds problematic, please let us know how you’re using the data (which languages and entity types) and how much time you would need to support Wikidata.

Editing: Making MusicBrainz better

Over the past few weeks I’ve received a number of emails from people who are concerned about some editors who are losing sight of some basic principles behind editing data in MusicBrainz. I wanted to chime in and remind people of some of the principles that should guide how we all get along when we edit data in MusicBrainz.

First and foremost is:

Be polite and give people the benefit of the doubt that they are doing the right thing.

I don’t have to explain being polite. Yes, we all have our bad days — that is a given. But if you’re having a bad day, stop editing MusicBrainz and step away from your computer. Go outside! When you do edit, please be kind to your fellow editors.

Giving people the benefit of the doubt that they are doing the right thing is also important. The vast majority of people who edit MusicBrainz have good intentions and you should assume that to be the case.

Second, edit to make the database better. Vote yes if an edit makes the data better.

This one is a lot more vague, since “better” is a subjective term. We should accept edits that are “good enough” and avoid asking people to make “perfect” edits.

Edits fit into four categories:

  1. Edits that makes things better (perfect or not)
  2. Edits makes things different (but neither are better)
  3. Edits that contain some correct things and some incorrect things
  4. Edits that are outright wrong (existing data is better)

The first type should clearly get a yes vote. For the second, if it doesn’t make things worse, abstain and leave a comment. The third is a judgement call and I would suggest applying this heuristic:

Unless it takes more time to fix the edit than to make a new one, vote yes.

Clearly, the fourth type deserves a no vote.

That brings me to the final topic for now: No votes. A no vote is a very strong expression that has potentially chilling effects that may prevent people from editing again. A no vote should be considered the last resort. Use a no vote if you can’t find another way to resolve an edit.

Finally, some tips for auto editors: If you see an edit that is not perfect, approve it and fix it.

Auto editors are supposed to set the tone for the project and auto editors should practically never vote no on something. You have more powers than fellow editors, so please use your powers for good!

Thanks and happy (and polite) editing!

Announcing the AcousticBrainz project

MetaBrainz and the Music Technology Group at Universitat Pompeu Fabra are pleased to announce the first public release of the AcousticBrainz project.

What is AcousticBrainz?
The AcousticBrainz project aims to crowd source acoustic information for all of the music in the world and make it available to the public. The goal of AcousticBrainz is to provide music technology researchers and open source hackers with a massive database of information about music.

AcousticBrainz uses a state of the art research project called Essentia (, developed over the last 10 years at the Music Technology Group.

Data generated from processing audio files with Essentia is collected by the AcousticBrainz project and made available to the public under the CC0 license (public domain). In 6 weeks since its inception, AcousticBrainz contributors have already submitted data for 650,000 audio tracks using pre-release software.

Today we are releasing client programs to submit data to the AcousticBrainz server and our first public release containing audio features for over 650,000 audio files.

What data does it have?
AcousticBrainz contains information called audio features. This acoustic information describes the acoustic characteristics of music and includes low-level spectral information such as tempo, and additional high level descriptors for genres, moods, keys, scales and much more. These features are explained in more detail at

How can I get it?
You can access AcousticBrainz data via our API. See details at
We also provide downloadable dumps of the whole dataset. You can download it (all 13 gigabytes!) at

What can I do with it?
We hope that this database will spur the development of new music technology research and allow music hackers to create new and interesting recommendation and music discovery engines. Here are some ideas of things we would like to see:

  • Music discovery
  • Playlist generation
  • Improving the state of the art in genre recognition
  • Analytics on the musical structure of popular music
  • and more!

This is one of the largest datasets of this kind available for research, and the only one of this size that we know of which contains both freely available data as well as the reference source code used to compute the data.

How can I contribute?
If you are a music researcher, you can help us by contributing to the essentia project. Go to the essentia homepage to see how you can do this. If you do something cool with the data let us know. We’d like to start a “made with AcousticBrainz” page where we will showcase interesting projects.

If you have any audio files, we would love for you to contribute audio features to our project. You can do this by downloading our submission clients from We provide clients for Windows, Mac, and Linux.

If you find any bugs or errors in the AcousticBrainz stack please let us know! Report issues to

We can’t wait to see what kind of things you will make with our data.

The AcousticBrainz team.

May 2014 virtual machines released

I’m pleased to announce that we’ve just released the May 2014 Virtual Machine images of the MusicBrainz server!

If you’d like to direct download or BitTorrent download these images, head on over to our Server Setup page.

There have been no real changes to the setup of the VM images — we still publish a VirtualBox and a VMWare image. The instructions for using the VMs haven’t changed; the latest data and the May 26th version of the software are loaded on these images.

Have fun!

LinkedBrainz: Alive and well!

Barry Norton has been a star and has created and hosted RDF dumps of the MusicBrainz data and also established a permanent SPARQL endpoint for our data on

The timing of this is perfect, because our next release will remove the RDFa from our pages. Proper RDF data and a SPARQL end-point are the best ways to move forward with MusicBrainz data in the context of linked open data.

This gives the MusicBrainz development team the freedom to focus on making MusicBrainz better while leaving the nitty gritty parts of making our data friendly to the linked open data hackers to experts like Barry.

Thanks so much for making this happen, Barry!

Important Update on Replication

We’ve been informed that some people have noticed their MusicBrainz slave servers have stopped applying replication packets. We’ve tracked the problem down and now have a fix for this problem.

How Do I Tell If I’m Affected?

To determine if you are affected, connect to the MusicBrainz PostgreSQL database (we recommend using ./admin/psql from your musicbrainz-server checkout) and run the following:

SELECT now() - last_replication_date FROM replication_control;

If you see a date that is larger than a few days, there’s a good chance you are affected by this bug and should upgrade.

How Do I Fix This?

Fixing this is easy, you simply need to update your Git checkout. We suggest the following:

cd musicbrainz-server
git fetch --tags origin
git checkout v-2013-10-14-replication-fix-2

EDIT: formerly used a different (broken) tag and recommended a different method of fetching tags.

When replication next runs, the data that cannot be applied will be rewritten and replication will continue as normal.

Musing: compacting replication packets

I’ve recently been working to write more about MusicBrainz internals, and thoughts about the project. Often this blog doesn’t see many posts, and most of them are on official topics like releases, so putting this here is an experiment. I hope you’ll all enjoy hearing about something a bit less concrete (and perhaps less dry, more technical) than usual!

How Replication Works

Replication is a pretty important part of MusicBrainz, though perhaps to the average user it’s a bit hidden. For those readers who aren’t familiar with it, replication is the mechanism behind the live data feed: users have a PostgreSQL database and using tools we provide (or some third-party alternatives) regularly download and apply so-called “replication packets” which describe the changes to the database in a specific period.

Replication packets are .tar.bz2 archives with a collection of files in them: COPYING with the license info, README with a very sparse description of replication, SCHEMA_SEQUENCE with the version of the database schema the replication packet applies to, REPLICATION_SEQUENCE with a sequence number that the code uses to apply replication packets in the correct order, TIMESTAMP with, well, a timestamp, and finally a folder mbdump, which contains two files: dbmirror_pending and dbmirror_pendingdata. Those of you who use the MusicBrainz data dumps may recognize this format: it’s the same as the data dumps. dbmirror_pending and dbmirror_pendingdata are two database tables that are used by replication to store the data about changes while those changes are being applied to the database.

Let’s look more closely at what those two tables contain. dbmirror_pending has columns for a sequence ID, a fully-qualified table name, an operation, and a transaction ID. dbmirror_pendingdata has columns for a sequence ID, a boolean for if data specifies keys only, and finally for the change data itself. Jointly, these two tables combine, conceptually, into an ordered list of operations to perform on the database. Since I tend to think in JSON, here’s a way you could imagine a single operation looking:

{"table": "musicbrainz.release",
"operation": "update",
"existing_row": "\"id\"='1290306' \"name\"='620308' \"artist_credit\"='234861' \"release_group\"='1269028' \"status\"='1' \"packaging\"='1' \"country\"='150' \"language\"='120' \"script\"=",
"update_row": "\"id\"='1290306' \"gid\"='e37dfeea-0f25-48fa-85c0-b4d174ff172d' \"name\"='620308' \"artist_credit\"='234861' \"release_group\"='1269028' \"status\"='1' \"packaging\"='1' \"country\"='150' \"language\"='120' \"script\"= \"date_year\"='2009' \"date_month\"= \"date_day\"= \"barcode\"='8715777007870' \"comment\"='' \"edits_pending\"='3' \"quality\"='-1' \"last_updated\"='2013-05-15 13:01:05.065623+00'"}

This operation would specify that it should update the table musicbrainz.release by taking the row whose id is 1290306, gid is e37dfeea-0f25-48fa-85c0-b4d174ff172d, name is 620308, etc. (as listed in ‘existing_row’) and change it to have id 1290306, gid e37dfeea-0f25-48fa-85c0-b4d174ff172d, name 620308, etc. (as listed in update_row).

Compacting replication packets

So there’s a summary of how replication works, in rough, abstract terms. Now on to the real topic of the post: making replication packets more compact. As the current system works, every change is put into the replication packet; that is, if in the course of an hour (one replication packet), a table changes twenty times, then twenty operations will end up in the replication packet. Sometimes this is useful: some data users use database-level triggers to update their own derived information, and sometimes that requires seeing every change, even if it changes again very soon thereafter. However, for most people, replication is just a way to get their database up to date every hour. For these people, having all twenty updates is wasteful — as far as they’re concerned, it could just be a simple update from how the row looked at the start of the hour to how it looked at the end. This becomes especially true for the (currently rather underused) larger replication packets (daily and weekly, specifically), which include more changes (and thus more changes to the same rows).

To formalize this idea a bit more, let’s make one more abstraction: a chain of operations. A chain is, informally, an ordered list of operations on the same data (or the same row). The simplest chain is just one operation, and from there we can work with chains in a variety of ways:

  • Chains can be combined: if the final state of a chain corresponds to the initial state of another chain, and the first chain’s final operation is before the initial operation of the second (without any other chains whose initial state corresponds to the first chain’s final state in-between), then the two chains can be combined into one chain.
  • Chains can be reordered: if there is no way to combine two chains by the above rule (including by way of intermediate operations or chains), then the two chains operate on completely separate data and thus can happen in any order. (For the database-savvy who might have noticed: slave databases, that use replication, don’t have foreign keys or other constraints which might make this not true).
  • Chains can be combined even more: If the final operation of a chain is a deletion, and the initial operation of another is an insertion, the two chains can be combined by turning the deletion + insertion into a single update. (As you might imagine, the ability to change the order of chains helps a lot here!)
  • Chains can be collapsed: perhaps most important! Any number of updates in a chain can be turned into a single update, from the initial state of the first update to the final state of the last update. Additionally, an insertion followed by an update can be turned into a single insertion, directly to the final state of the update. Additionally, an update followed by a deletion can be turned into a single deletion, directly from the initial state of the update. Finally, an insertion followed by a deletion can be ignored entirely, since it has no lasting effect on the database.

Thus, by creating, combining, reordering, and collapsing chains, we can make replication packets do many fewer operations (which, in turn, can make the packets smaller and have them apply faster). I’ve still glossed over the details of how this could be implemented, though, so to wrap things up, here’s a basic algorithm for optimizing packets:

  1. Loop through operations in the order ProcessReplicationChanges (the script responsible for applying changes to the database from a replication packet) would: order the transactions in ascending order by the maximum sequence ID within the transaction, and within a transaction by ascending sequence ID.
  2. Take action depending on the type of operation:
    • If an insertion: see if the output already contains a deletion on the same table. If so, take the initial state of the deletion and the final state of the insertion and output an update instead of the deletion. If no such deletion exists, simply copy the insertion to the output.
    • If an update: see if the output already contains an insertion or an update on the same data (that is, find an operation whose final state matches the initial state of the update). If you find one, replace it with an operation of the same type (and initial state, if applicable) but with the final state of the new update instead of what it previously included. If you don’t find one, just copy over the update to the output.
    • If a deletion: see if the output already contains an insertion on the same data. If it does, remove it, add nothing to the output, and move on. If not, look for an update on the same data. If you find one, replace it with a deletion from the initial state of the update. If not, look for an insertion on the same table. If you find one, replace it with an update from the initial state of the deletion to the final state of the insertion. Finally, if you found nothing, copy the deletion to the output.
  3. Dump the output as a replication packet. Transactions shouldn’t matter, so put them all in the same transaction.

Final notes, FAQs

  • Why isn’t this already being done? Daily and weekly packets are relatively new, and since some data users do want to see every single operation, it doesn’t make sense to do this to the hourly packets. It’s also somewhat annoying to get the operations in the right order, because they aren’t in the replication packet dump, due to how transactions have to be processed. The musicbrainz-server code gets around this by importing the data to a PostgreSQL table and letting the database do the work of putting things in the right order. mbslave instead loads the entire packet into memory and sorts it there, which obviously is potentially dangerous with a larger packet (which, as noted above, are more likely to derive value from this process). Altogether, the safest way to implement this would be to mimic musicbrainz-server’s process, but to do this on the production servers it would need to use different table names so as to not interfere with the normal replication packet creation process. But mostly, because nobody’s written it yet.
  • How much benefit would this really bring? Simple answer: I don’t really know, but I know it’s some. Autoedits (additions, especially) have a tendency to produce multiple operations where one would do fine, because they first increment the ‘edits_pending’ column of whatever they’re editing in one operation, then in the process of applying the edit (automatically, and immediately, in the same transaction) decrement it. Editors also tend to change a bunch of different things about the same entity all at once, but sometimes not in one edit but rather in several. Of course, any deletion is easily counterbalanced by the many insertions that happen all the time in MusicBrainz; perhaps most notably, daily scripts run to clean up unused entities, so any packet that includes those changes will probably be able to collapse some insertion + deletion pairs. So there’s several cases where obvious chains would appear already.

Hopefully those of you who’ve made it down this far found this enlightening — or, at least, interesting. At some point in the future this might be something we do with the daily and weekly replication packets we’re already creating (currently just by concatenating together hourly packets), but for now there’s no solid plan to do so. Thus, just musing for now!

Thanks for reading!