Ayende @ Rahien

Oren Eini aka Ayende Rahien CEO of Hibernating Rhinos LTD, which develops RavenDB, a NoSQL Open Source Document Database.

You can reach me by:

oren@ravendb.net

+972 52-548-6969

Posts: 7,074 | Comments: 49,800

filter by tags archive
time to read 2 min | 297 words

imageI mentioned in my previous post that I managed to lock myself out of the car by inputting the wrong pin code. I had to wait until the system reset before I could enter the right pin code. I got a comment to the post saying that it would be better to use a thumbprint scanner for the task, to avoid this issue.

I couldn’t disagree more.

Let’s leave aside the issue of biometrics, their security and the issue of using that for identity. I don’t want to talk about that subject. I’ll assume that biometrics cannot fail and can 100% identify a person with no mistakes and no false positives and negatives.

What is the problem with a thumbprint vs. a pin code as the locking mechanism on a car?

Well, what about when I need someone else to drive my car? The simplest example may be valet parking, leaving the car at the shop or just loaning it to someone.  I can give them the pin code over the phone, I’m hardly going to mail someone my thumb because. There are many scenarios where I actually want to grant someone the ability to drive my car, and making it harder to do so it a bad idea.

There is also the issue of what happens if my thumb is inoperable? It might be raining and my hands are wet, or I changed a tire and will need half an hour at the sink to get cleaned up again.

You can think up solutions to those issues, sure, but they are cases where the advanced solution makes anything out of the ordinary a whole lot more complex. You don’t want to go there.

time to read 3 min | 582 words

Complex systems are considered problematic, but I’m using this term explicitly here. For reference, see this wonderful treatise on the subject. Another way to phrase this is how to create robust systems with multiple layers of defense against failure.

When something is important, you should prepare in advance for it. And you should prepare for failure. One of the typical (and effective) methods for handle failures is the good old retry. I managed to lock myself out of my car recently, had to wait 15 minutes before it was okay for me to start it up. But a retry isn’t going to help you if the car run out of fuel, or there is a flat tire. In some cases, a retry is just going to give you the same failing response.

Robust systems do not have a single way to handle a process, they have multiple, often overlapping, manners to do their task. There is absolutely duplication between these methods, which tend to raise the hackles of many developers. Code duplication is bad, right? Not when it serve a (very important) purpose.

Let’s take a simple example, order processing. Consider the following example, an order made on a website, which needs to be processed in a backend system:

image

The way it would work, the application would send the order to a payment provider (such as PayPal) which would process the actual order and then submit the order information via web hook to the backend system for processing.

In most such systems, if the payment provider is unable to contact the backend system, there will be some form of retry. Usually with a exponential back-off strategy. That is sufficient to handle > 95% of the cases without issue. Things get more interesting when we have a scenario where this breaks. In the above case, assume that there is a network issue that prevent the payment provider from accessing the backend system. For example, a misconfigured DNS entry means that external access to the backend system is broken. Retrying in this case won’t help.

For scenarios like this, we have another component in the system that handles order processing. Every few minutes, the backend system queries the payment provider and check for recent orders. It then processes the order as usual. Note that this means that you have to handle scenarios such as an order notification from the backend pull process concurrently with the web hook execution. But you need to handle that anyway (retrying a slow web hook can cause the same situation).

There is additional complexity and duplication in the system in this situation, because we don’t have a single way to do something.

On the other hand, this system is also robust on the other direction. Let’s assume that the backend credentials for the payment provider has expired. We aren’t stopped from processing orders, we still have the web hook to cover for us.  In fact, both pieces of the system are individually redundant. In practice, the web hook is used to speed up common order processing time, with the backup pulling recent orders as backup and recovery mechanism.

In other words, it isn’t code duplication, it is building redundancy into the system.

Again, I would strongly recommend reading: Short Treatise on the Nature of Failure; How Failure is Evaluated; How Failure is Attributed to Proximate Cause; and the Resulting New Understanding of Patient Safety.

time to read 5 min | 970 words

A customer reported that on their system, they suffered from frequent cluster elections in some cases. That is usually an indication that the system resources are hit in some manner. From experience, that usually means that the I/O on the machine is capped (very common in the cloud) or that there is some network issue.

The customer was able to rule these issues out. The latency to the storage was typically withing less than a millisecond and the maximum latency never exceed 5 ms. The network monitoring showed that everything was fine, as well. The CPU was hovering around the 7% CPU and there was no reason for the issue.

Looking at the logs, we saw very suspicious gaps in the servers activity, but with no good reason for them. Furthermore, the customer was able to narrow the issue down to a single scenario. Resetting the indexes would cause the cluster state to become unstable. And it did so with very high frequency.

“That is flat out impossible”, I said. And I meant it. Indexing workload is something that we have a lot of experience managing and in RavenDB 4.0 we have made some major changes to our indexing structure to handle this scenario. In particular, this meant that indexing:

  • Will run in dedicated threads.
  • Are scoped to run outside certain cores, to leave the OS capacity to run other tasks.
  • Self monitor and know when to wind down to avoid impacting system performance.
  • Indexing threads are run with lower priority.
  • The cluster state, on the other hand, is run with high priority.

The entire thing didn’t make sense. However… the customer did a great job in setting up an environment where they could show us: Click on the reset button, and the cluster become unstable.  So it is impossible, but it happens.

We explored a lot of stuff around this issue. The machine is big and running multiple NUMA node, maybe it was some interaction with that? It was highly unlikely, and eventually didn’t pan out, but that is one example of the things that we tried.

We setup a similar cluster on our end and gave it 10x more load than what the customer did, on a set of machines that had a fraction of the customer’s power. The cluster and the system state remain absolutely stable.

I officially declared that we were in a state of perplexation.

When we run the customer’s own scenario on our system, we saw something, but nothing like what we saw on their end. One of the things that we usually do when we investigate resource constraint issues is to give the machines under test a lot less capability. Less memory and slower disks, for example, means that it is much easier to surface many problems. But the worse we made the situation for the test cluster, the better the results became.

We changed things up. We gave the cluster machines with 128 GB of RAM and fast disks and tried it again. The situation immediately reproduced.

Cue facepalm sound here.

Why would giving more resources to the system cause instability in the cluster? Note that the other metrics also suffered, which made absolutely no sense.

We started digging deeper and we found the following index:

It is about as simple an index as you can imagine it would be and should cause absolutely no issue for RavenDB. So what was going on? We then looked at the documents…

image

I would expect the State field to be a simple enum property. But it is an array that describe the state changes in the system. This array also holds complex objects. The size of the array is on average about 450 items and I saw it hit a max of 13,000 items.

That help clarify things. During index, we have to process the State property, and because it is an array, we index each of the elements inside it. That means that for each document, we’ll index between 400 – 13,000 items for the State field. What is more, we have a complex object to index. RavenDB will index that as a JSON string, so effectively the indexing would generate a lot of strings. These strings are going to be held in memory until the end of the indexing batch. So far, so good, but the problem in this case was that there were enough resources to have a big batch of documents.

That means that we would generate over 64 million string objects in one of those batches.

Enter GC, stage left.

The GC will be invoked based on how many allocations you have (in this case, a lot) and how many live objects you have. In this case, also a lot, until the index batch is completed. However, because we run GC multiple times during the indexing batch, we had promoted significant numbers of objects to the next generation, and Gen1 or Gen2 collections are far more expensive.

Once we knew what the problem was, it was easy to find a fix. Don’t index the State field. Given that the values that were indexed were JSON strings, it is unlikely that the customer actually queried on them (later confirmed by talking to the customer).

On the RavenDB side, we added monitoring for the allocation frequency and will close the indexing batch early to prevent handing the GC too much work all at once.

The reason we failed to reproduce that on lower end machine was simple, RavenDB already used enough memory so we closed the batch early, before we could gather enough objects to cause the GC to really work hard. When running on a big machine, it had time to get the ball rolling and hand the whole big mess to the GC for cleanup.

time to read 1 min | 65 words

I had the pleasure of talking with Vaishali Lambe in SoLeadSaturday this week.

We talked about various aspects of being in - Building a business around open source software, Working in a distributed teams, Growing a company from 1 employee to 30+, Non technical details that are important to understand to run a company. Also, discussed an importance of blogging.

You can check it here.

time to read 3 min | 541 words

One of the distinguishing feature of RavenDB is its ability to process large aggregations very quickly. You can ask questions on very large data sets and get the results in milliseconds. This is interesting, because RavenDB isn’t an OLAP database and the kind of questions that we ask can be quite complex.

For example, we have the Products/Recommendations index, which allow us to ask:

For any particular product, find me how many times it was sold, what other products were sold with it and in what frequency.

The index to manage this is here:

The way it works, we map the orders and have a projection for each product, and then we add the other products that were sold with the current one. In the reduce, we group by the product and aggregate the related products together.

But I’m not here to talk about the recommendation engine. I wanted to explain how RavenDB process such indexes. All the information that I’m talking about can be seen in the Map/Reduce visualizer in the RavenDB Studio.

Here is a single entry for this index. You can see that products/11-A was sold 544 times and 108 times with products/69-A.

image

Because of the way RavenDB process Map/Reduce indexes, when we query, we run over the already precomputed results and there is very little computation cost at querying time.

Let see how RavenDB builds the index. Here is a single order, where three products were sold. You can see that each of them as a very interesting tree structure.

image

Here is how it looks like when we zoom into a particular product. You can see how RavenDB aggregate the data. First in the bottom most page on the right (#596). We aggregate that with the other 367 pages and get intermediate results at page #1410. We then aggregate that again with the intermediate results in page #105127 to get the final tall. In this case, you can see that products/11-A was sold 217,638 times and mostly with products/16-A (30,603 times) and products/72-A (20,603 times).

image

When we have a new order, all we need to do is update a bottom most page and then recurse upward in the three. In the case we have here, there is a pretty big reduce value and we are dealing with tens of millions of orders. We have three levels to the tree, which means that we’ll need to do three update operations to account for new or updated data. That is cheap, because it means that we have to do very little work to maintain the index.

At query time, of course, we don’t really have to do much, all the hard work was done.

I like this example, because it shows case a non trivial example and how RavenDB handles this with ease. These kind of non trivial work is something that tend to be very hard to get working properly and with RavenDB this is part of my default: “let’s do this on the fly demo”.

time to read 3 min | 452 words

It should come as no surprise that our entire internal infrastructure is running on RavenDB. I wholly believe in the concept of dog fooding and it has serve us very well over the years.

I was speaking to a colleague just now and it occurred to me that it is surprising that we do certain things wrong, intentionally. It is fair to say that we know what the best practices for using RavenDB are, the things that you can do to get the most out of it.

In some of our internal systems, we are doing things in exactly the wrong way. We are doing things that are inefficient in RavenDB. We take the expedient route to implement things.  A good example of that is that we have a set of documents that can grow to be multiple MB in size. They are also some of the most common changed documents in the system. Properly design would call to break them apart to make things easier for RavenDB.

We intentionally modeled things this way. Well, I gave the modeling task to an intern with no knowledge of RavenDB and then I made things worse for RavenDB in a few cases where he didn’t get it out of shape enough for my needs.

Huh?! I can hear you thinking. Why on earth would we do something like that?

We do this because if serves as an excellent proving ground for misuse of RavenDB. It show us how the system behave under non ideal situations. Not just when the user is able to match everything to the way RavenDB would like things to be, but how they are likely to build their system. Unaware of what is going on behind the scenes and what the optimal solution would be. We want RavenDB to be able to handle that scenario well.

An example that pops to mind was having all the uploads on the system be attachments on a single document. That surfaced that we had a O(N^2) algorithm very deep in the bowels of RavenDB for placing a new attachment. It would be completely invisible under normal case, because it was fast enough under any normal or abnormal situation that we could think of. But when we started getting high latency from uploads, we realized that adding the 100,002th attachment to a document required us to scan through the whole list… it was obvious that we needed a fix. (And please, don’t put hundreds of thousands of attachments on a document, it will work (and it is fast now), but it isn’t nice).

Doing the wrong thing on purpose means that we can be sure that when users are doing the wrong thing accidently, they get good behavior.

time to read 3 min | 577 words

email-me-clipart | free clip art from: www.fg-a.com/email1.s… | FlickrWe got a feature request that we don’t intend to implement, but I thought the reasoning is interesting enough for a blog post. The feature request:

If there is a critical error or major issue with the current state of the database, for instance when the data is not replicated from Node C to Node A due to some errors in the database or network it should send out mail to the administrator to investigate on  the issue. Another example is, if the database not active due to some errors then it should send out mail as well.

On its face, the request is very reasonable. If there is an error, we want to let the administrator know about it, not hide it in some log file. Indeed, RavenDB has the concept of alerts just for that reason, to surface any issues directly to the admin ahead of time. We also have a mechanism in place to allow for alerts for the admin without checking in with the RavenDB Studio manually: SNMP. The Simple Network Monitoring Protocol is designed specifically to enable this kind of monitoring and RavenDB expose a lot of state via that which you can act upon in your monitoring system.

Inside your monitoring system, you can define rules that will alert you. Send an SMS if the disk space is low, or email on an alert from RavenDB, etc. The idea of actively alerting the administrator is something that you absolutely want to have.

Having RavenDB send those emails, not so much. RavenDB expose monitoring endpoint and alerts, it doesn’t act or report on them. That is the role of your actual monitoring system. You can setup Zabbix or talk to your Ops team which likely already have one installed.

Let’s talk about the reason that RavenDB isn’t a monitoring system.

Sending email is actually really hard. What sort of email provider do you use? What options are required to set it up a connection? Do you need X509 certificate or user/pass combo? What happens if we can’t send the email? That is leaving aside the fact that actually getting the email delivered is hard enough. Spam, SPF, DKIM and DMARC is where things start. In short, that is a lot of complications that we’ll have to deal with.

For that matter, what about SMS integration? Surely that would also help. But no one uses SMS today, we want WhatsApp integration, and Telegram, and … You go the point.

Then there are social issues. How will we decide if we need to send an email or not? There should be some policy, and ways to configure that. If we won’t have that, we’ll end up sending either too many emails (which will get flagged / ignored) or too few (why aren’t you telling me about XYZ issue?).

A monitoring system is built to handle those sort of issues, it is able to aggregate reports and give you a single email with the current status, open issues for you to fix and do a whole lot more that is simply outside the purview or RavenDB. There is also the most critical alert of all, if RavenDB is down, it will not be able report that it is down because it is down.

The proper way to handle this is to setup integration with a monitoring system, so we’ll not be implementing this feature request.

time to read 6 min | 1084 words

In my previous post, I wrote about the case of a medical provider that has a cloud database to store its data, as well as a whole bunch of doctors making house calls. There is the need to have the doctors have (some) information on their machine as well as push updates they make locally back to the cloud.

image

However, given that their machines are in the field, and that we may encounter a malicious doctor, we aren’t going to fully trust these systems. We still want the system to function, though. The question is how will we do it?

Let’s try to state the problem in more technical terms:

  • The doctor need to pull data from the cloud (list of patients to visits, patient records, pharmacies and drugs available, etc).
  • The doctor nee to be able to create patient records (exam made, checkup results, prescriptions, recommendations, etc).
  • The doctor’s records needs to be pushed to the cloud.
  • The doctor should not be able to see any record that is not explicitly made available to them.
  • The same applies for documents, attachments, time series, counters, revisions, etc.

Enforcing Distributed Data Integrity

The requirements are quite clear, but they do bring up a bit of a bother. How are we going to enforce it?

One way to do that would be to add some metadata rule for the document, deciding if a doctor should or should not see that document. Something like that:

image

In this model, a doctor will have be able to get this document if they have any of the tags associated with the document.

This can work, but that has a bunch of non trivial problems and a huge problem that may not be obvious. Let’s start from the non trivial issues:

  • How do you handle non document data? Based on the owner document, probably. But that means that we have to have a parent document. That isn’t always the case.
  • It isn’t always the case if the document was deleted, or is in a conflicted state.
  • What do you do with revisions, if the access tags has changed? What do you follow?

There are other issues, but as you can imagine, they are all around managing the fact that this model allows you to change the tags for the document and expect to handle this properly.

The huge problem, however, is what should happen when a tag is removed? Let’s assume that we have the following sequence of events:

  • patients/oren is created, with access tag of “doctors/abc”
  • That access tag is then removed
  • Doctor ABC’s machine is then connected to the cloud and setup replication.
  • We need to remove patients/oren from the machine, so we send a tombstone.

So far, so good. However, what about Doctor' XYZ’s machine? At this time, we don’t know what the old tags were, and that machine may or may not have that document. It shouldn’t have it now, so we’ll send a tombstone there? That leads to information leak by revealing document ids that we aren’t authorized for.

We need a better option.

Using the Document ID as the Basis for Data Replication

We can define that once created, the access tags are immutable, and that would help considerably.  But that is still fairly complex to manage and opens up issues regarding conflicts, deletion and re-creation of a document, etc.

Instead, we are going to use the document’s id as the source for the decision to replicate the document or not. In other words, when we register the doctor’s machine, we set it up so it will allow:

Incoming paths Outgoing paths
  • doctors/abc/visits/*
  • tasks/doctors/abc/*
  • patients/clinics/33-conventry-rd/*
  • pharmacies/*
  • tasks/doctors/abc/*
  • doctors/abc
  • laboratories/*

In this case, incoming and outgoing are defined from the point of view of the cloud cluster. So this setup allows the doctor’s machine to push updates to any document with an id that starts with “doctors/abc/visits/” or “tasks/doctors/abc/*”. And the cloud will send all pharmacies and laboratories data. The cloud will also send all the patients for the doctor’s clinic as well as the tasks for this doctor, finally, we have the doctor’s record itself. Everything else will be filtered.

This Model is Simple

This model is simple, it provides a list of outgoing and incoming paths for the data that will be replicated. It is also quite surprisingly powerful. Consider the implications of the configuration above.

The doctor’s machine will have a list of laboratories and pharmacies (public information) locally. It will have the doctor’s own document as well as records of the patients in the clinic. The doctor is able to create and push patient visit’s records. Most interestingly, the tasks for the doctor are defined to allow both push and pull. The doctor will receive updates from the office about new tasks (home visits) to make and can mark them complete and have it show up in the cloud.

The doctor’s machine (and the doctor as well) is not trusted. So we limit the exposure of the data that they can see on a Need To Know basis. On the other hand, they are limited in what they can push back to the cloud. Even with these limitations, there is a lot of freedom in the system, because once you have this defined, you can write your application on the cloud side and on the laptop and just let RavenDB handle the synchronization between them. The doctor doesn’t need access to a network to be able to work, since they have a RavenDB instance running locally and the cloud instance will sync up once there is any connectivity.

We are left with one issue, though. Note that the doctor can get the patients’ files, but is unable to push updates to them. How is that going to work?

The reason that the doctor is unable to write to the patients’ files is that they are not trusted. Instead, they will send a visit record, which contains their finding and recommendation and on the cloud, we’ll validate the data, merge it with the actual patient’s record, apply any business rules and then update the record. Once that is done, it will show up in the doctor’s machine magically. In other words, this setup is meant for untrusted input.

There are more details that we can get into, but I hope that this outline the concepts clearly. This is not a RavenDB 5.0 feature, but will be part of the next RavenDB release, due around September.

time to read 4 min | 753 words

RavenDB is typically deployed as a set of trusted servers. The network is considered to be hostile, which is why encrypt everything over the wire and using X509 certificates for mutual authentication, but once the connection is established, we trust the other side to follow the same rules as we do.

To clarify, I’m talking here about trust between nodes, not a client connected to RavenDB. These are also authenticated using X509 certificate, but they are limited to the access permissions assigned to them. Nodes in a cluster fully trust one another and need to do things like forward commands accepted by one node to another one. That requires that the second node trust that the first node properly authenticated the client and won’t pass operations that the client has no authority for.

Use Case 1: A Database System for Multiple Medical Clinics

I think that a real use case might make things more concrete. Let’s assume that we have a set of clinics, with the following distribution of data.

image 

We have two clinics, one in Boston and one in Chicago, as well as a cloud system. The rules of the system are as follows:

  • Data from each clinic is replicated to the cloud.
  • Data from the cloud is replicated to the clinics.
  • Data from a clinic may only be at the clinic or in the cloud.
  • A clinic cannot get (or modify) data that didn’t came from the clinic.

In this model, we have three distinct locations, and we presumably trust all of them (otherwise, would we put patient data on them?). There is a need to ensure that we don’t expose patient data from one clinic to another, but that is about it. Note that in terms of RavenDB topology, we don’t have a single cluster here. That wouldn’t make sense. To start with, we need to be able to operate the clinic when there is no internet connectivity. And we don’t want to pay with any avoidable latency even if everything is working fine.  So in this case, we have three separate clusters, one in each location, and they are connected to one another using RavenDB’s multi master replication.

Use Case 2: A Database System Sharing with Outside Edge Points

Let’s look at another model, however, in this case, we are still dealing with medical data, but instead of a clinic, we have to deal with a doctor making house calls:

image

In this case, we are still talking about private data, but we are no longer trusting the end device. The doctor may lose the laptop, they may have malware running on the machine or may be trying to do Bad Things directly.  We want to be able to push data to the doctor’s machine and receive updates from the field.

RavenDB has some measures at the moment to handle this scenario. You need to only get some data from the cloud to the doctor’s laptop, and you want to push only certain things back to the cloud. You can use pull replication and ETL. to handle this scenario, and it will work, as long as you are willing to trust the end machine. Given the stringent requirement for medical data, it is not something out of bounds, actually. Full volume encryption, forbidding use of unknown software and a few other protections ensure that if the laptop is lost, the only thing you can do with it is repurpose the hardware.  If we can go with that assumption, this is great.

However… we need to consider the case that our doctor is actually malicious.

image

When the Edge Point isn’t as Healthy as the Doctor Using It

So we need a something in the middle, between all our data and what can reside on that doctor’s machine.  As it currently stands, in order to create the appropriate barrier between the doctor’s machine and the cloud, you’ll have to write your own sync code and apply any logic / authorization at that level.

Sync code is non trivial, mostly because of the number of edge cases you have to deal with and the potential for conflicts. This has already been solved by RavenDB, so having to write it again is not ideal as far as we are concerned.

What would you do?

time to read 2 min | 380 words

A sadly common place “attack” on applications is called “Web Parameter Tampering”. This is the case where you have a URL such as this:

https://secret.medical.info/?userid=823

And your users “hack” you using:

https://secret.medical.info/?userid=999

And get access to another users records.

As an aside, that might actually be considered to be hacking, legally speaking. Which make me want to smash my head on the keyboard a few time.

Obviously, you need to run your security validation on parameters, but there are other reasons to want to avoid to expose the raw identifiers to the user. If you are using the a incrementing counter of some kind, creating two values might cause you to leak the rate in which your data change. For example, a competitor might want to create an order once a week and track the number of the order. That will give you a good indications of how many orders there have been in that time frame.

Finally, there are other data leakage issues that you want to might want to take into account. For example, “users/321” means that you are likely to be using RavenDB while “users/4383-B” means that you are using RavenDB 4.0 or higher and “607d1f85bcf86cd799439011” means that you are using MongoDB.

A common reaction to this is to switch your ids to use guids. I hate that option, it means that you are entering very unfriendly territory  for the application. Guids convey no information to the developers working with the system and they are hard to work with, from a humane point of view. They are also less nice for the database systemto work with.

A better alternative is to simply mask the information when it leaves your system. Here is the code to do so:

You can see that I’m actually using AES encryption to hide the data, and then encoding it in the Bitcoin format.

That means that an identifier such as "users/1123" will result in output such as this:

bPSPEZii22y5JwUibkQgUuXR3VHBDCbUhC343HBTnd1XMDFZMuok

The length of the identifier is larger, but not overly so and the id is even URL safe Smile. In addition to hiding the identifier itself, we also ensure that the users cannot muck about in the value. Any change to the value will result in an error to unmask it.

FUTURE POSTS

  1. Webinar Recording: Advanced Search Scenarios in RavenDB - 13 hours from now
  2. Putting JSON in a block chain? First decide what your JSON is… - about one day from now
  3. The failure of a computer you didn't even know existed - 3 days from now

There are posts all the way to Oct 30, 2020

RECENT SERIES

  1. Webinar recording (10):
    27 Aug 2020 - The App that Guarantees You're Going Out This Saturday Night
  2. re (27):
    27 Oct 2020 - Investigating query performance issue in RavenDB
  3. Reminder (10):
    25 Oct 2020 - Online RavenDB In Action Workshop tomorrow via NDC
  4. Podcast (3):
    17 Aug 2020 - #SoLeadSaturday with Oren Eini
  5. RavenDB Webinar (3):
    01 Jun 2020 - Polymorphism at Scale
View all series

RECENT COMMENTS

Syndication

Main feed Feed Stats
Comments feed   Comments Feed Stats