Next week I’ll be talking at CMU about how to build storage engines and what you need to consider when building them. The talk is open to the public (don’t have to be at CMU to be there).

Here are the details:

Event Date: Monday January 31, 2022
Event Time: 04:30pm EST
Location: https://cmu.zoom.us/j/95002789605?pwd=eEtJRnRaNnQ1bFZmTnUwbDRMaXBRQT09

In this talk, Oren Eini, founder of RavenDB, will discuss the design decisons and the manner in which RavenDB deals with storing data on disk.
Achieving highly concurrent and transactional system can be a challenging task. RavenDB solves this issue using a storage engine called Voron. We’ll go over the design of Voron and how it is able to achieve both high performance and maintain ACID integrity.

This talk is part of the Vaccination Database (Booster) Tech Talk Seminar Series.

Zoom Link: https://cmu.zoom.us/j/95002789605 (Passcode 982149)

Databases will typically divide their data files into pages, and manage all data in pages. As you can imagine, the notion of pages is pretty important to a database. Up to this point, we worked with pages, but not really. A page was just a buffer that we could read and write to. It took me 13 posts to manage to read and write pages from disk, but that is just the beginning. We need to add some structure for the pages. The usual way to do that is to have a page header, which contains some metadata about the page. That leads to an interesting question, how do we manage the page metadata?

One way to do that is to reserve a part of the page for this purpose, like so:

This is probably the simplest option (and what we use in Voron), but I don’t like it very much. It makes certain operations harder to deal with, because the usable size of the page is no longer a power of 2. Assuming we have 8KB pages and a page header of 64 bytes, that means that we can use just 8,128 bytes in the page. A lot of data structures are far easier to use if we can assume that the size is a power of 2.

Another thing to remember is that the underlying pager that we use operates on the level of 2 MB chunks. In other words, reading a value from the same 2MB range is as cheap as reading a value from the same 8KB page. For that reason, we can utilize the following approach, we’ll reserve the first section of the chunk for all the page headers. Here is what this will look like:

A chunk is 2MB in size, and pages are 8KB. That gives us 256 pages in a chunk. If we reserve the first page for page headers, that gives us 32 bytes for each page header. That is quite a lot, to be honest, which is great, because having more space on the header gives us a lot of options down the road.

An important consideration to pay attention to is that we read from the disk at 2MB chunks, but when we write, the writes happen at the page boundary. A change on Page #3, for example, means that we have to change both the header page and then page #3 as separate operations. If we were using an embedded page header, we’ll have just a single write to deal with. In practice, I don’t think this is a major issue. We are doing asynchronous writes and write coalescing anyway. For that matter, writes are actually done lazily (and the critical path goes to the Write Ahead Log anyway) so I don’t think this aspect matters so much.

At any rate, let’s get some idea about what the page header will look like:

I should note that this is merely a skeleton of the page header, just showing its structure. Over time, we’ll add a lot more details, but this is sufficient to understand the direction I’m going. A page header is denoted by a tag, which determines what type of page this is. Based on the type of the page, we can set various fields inside the header. You can also see the getPageCount(), which is a good demonstration of how we work with the page header. I need to have fine-grained control over the bytes in the header, so I’m setting them up in this manner.

Zig currently has a number of bugs related to packed / extern structs and unions, which make this a bit harder than expected, sadly. I’m waiting for the self hosted compiler, which will hopefully fix those issues.

With the PageHeader struct in place, we can start making use of this, like so:

There isn’t much going on here, which is great, but it does deserve some explanation. We first get the first page in the chunk for the page that was requested. For example, if we wanted to get page # 4, we get page #0, if we wanted to get page # 300, we’ll get page #256.  That page is the metadata page for the chunk, the only thing that it contains is the headers for all the pages in the chunk in question (including itself, of course).

The reason that this matters is that the header for the page is used to compute the number of pages that we need to return. As you’ll note, the API we have here does not provide a way for the caller to say how many pages should be returned. We need to figure it out on our own. To do that, however, we need to store that information somewhere. That is where the page header comes into place.

We get the page header of the page we care about from the metadata page of the chunk and compute the number of pages to get from the pager. Then we return to the caller the page header and the page buffer as a single unit.

Of course, this presupposes that the values got to the pages somehow. That leads us to other directions, however, dealing with page allocations. You can already see some hints of that in the code above, where we use freeSpaceBitmap. I’ll dedicate my next post to this topic.

In the last post, we figured out how we can find what pages we should flush to disk. In this one, I want to talk about how we can actually do the flushing. Once a page is marked as ready for flushing, we need to start an async process in which we write it to the disk. While that process is running, we need to ensure that there are no further writes happening on that page. We want to be able to handle flushing writes in the background with as little fuss as we possibly can, while at the same time maximizing parallelism. That turns out to be a significant challenge. Our previous write strategy simply gathered all the writes in memory and then flushed them to disk. Now we need to have an ongoing process. In addition, we also want to handle the scenario where we flush everything in memory for certain scenarios, such as checkpoints or orderly shutdown.

In short, the approach we previously took isn’t going to be sufficient for our needs. For that reason, I want to clear the board and start from scratch. As a reminder, here is the API that we provide:

The rules for this API are fairly simple. You acquire the writer and then may call the write() method any number of times. The writer will aggregate the writes in memory, and will use write behind to write the data to disk in the background. In other words, we aren’t necessarily going to write to the disk when write() is called, only to our own buffer.

Note that the writer is working using pages, not chunks, even though the buffer pool we use operates on chunks. In addition to the write() method, we have a flush() which pushes everything that is currently in the writer to the disk. That one is meant to be called when we want to create a checkpoint or when we do an orderly shutdown. The flush() call starts an asynchronous process that we can wait upon, so we aren’t going to stall on checkpoints.  While we write a page to disk, we’ll reject any writes to that page. This is important, because to do otherwise will mean that we may send messed up data to disk.

We also need to handle any and all sequences of writes. For example, consider:

• write(2, [1 page]);  // write a single page #2
• write(4, [4 pages]);// write 4 pages from page #4 (so pages: 4,5,6,7)

Now, we may actually get additional writes as well:

• write(6, [3 pages]);
• write(4, [1 page]);

In other words, there are no requirements from the caller about what should be called and in what order. The caller may write a set of pages and then repurpose them again. In this case, we expect that the end result is that we’ll issue the following writes to the disk:

• write(2, 1 page);
• write(4, 1 page);
• write(6, 3 pages);

Note that we have page #5, it was written in the first call to page #4 (when it used 4 pages), however, with the update of page #4 to be a single page, the data in page #5 is no longer relevant and we can skip writing it to disk. This is an effect of working with pages. The data doesn’t make sense outside of a page boundary, after all, so we should only write meaningful data.

After spending a lot of time thinking about it, I decided to go with the following behavior: The Pager’s writer will manage the write buffers, flushing to the disk when it is time. Callers can decide to let writes happen in the background or to wait for them. Here is the new state of the Writer:

There are a few interesting fields here:

• The lock is there to ensure that we have only a single writer. As a reminder, the actual I/O is asynchronous.
• The cache implements an Lru with a twist, I improved on that since the last post, more on this later.
• The parent field is the owning pager, obviously.
• The last three fields (inflight, completed and backgroundWrites) are all about handling the asynchronous writes for us.

Let’s take a look at how we handle the state for a particular write and then we’ll be able to discuss the implementation in depth:

On the surface, the WriteState is pretty simple. It contains the page we are writing and how many pages are involved in a particular write, but there are also quite a few additional fields that are… strange. The writer field is used when we are returning from a callback, the chunks are used to record which chunks are relevant for this particular write (remember, the Writer operates at the page level, while the Pager operates at the chunk level). Because we deal with individual pages, we have to ensure that each one of them has its own reference for the chunk in question. That also has the effect of pinning any dirty pages in memory until they are flushed to disk. The err field holds whether there was an error (and what was it) when writing to the disk. The list field is used in conjunction with the write behind policy and will be discussed shortly. We use the next field to create a single linked list with the head being the completed field on the Writer.

With that in mind, let’s look at how the write() method is implemented:

We start by draining any completed writes. It is important to detect any errors and to release the references from the chunks we locked during the write portion. We treat any error during the write process as fatal and abort immediately. At this point, we don’t know what the state of the system is. The only safe way to handle this is to shutdown and run recovery.

The next step is ensureNotWrittenTo() call. This is required because we may be in the process of writing a page to the disk, and we cannot modify the memory for that page (or pages) while they are being written to the disk. This is a recoverable error, by the way. The caller should wait until the pending writes are complete and retry.

We then check in the cache if the page is already there. If it is, we have to consider the potential for a mismatch in the sizes (described earlier in this post), when we have a previous write that spanned three pages but the current one only applies to a single page. Since the page boundary is important, we can just discard the previous write and continue. For the common case of a write to a page that was previously written to, however, we can simply copy the data from the provided buffer to the right page in memory and move on. Most of the hot pages should fit into this code path.

We then allocate the state for the current write, get the page (or pages) and copy the data from the provided buffer to our own memory. We also, crucially, push the new state into the writer’s cache. In the previous post that action would evict a page that we need to handle. In the current code, that is handled at the epilog of the method.

We iterate over the cache and evict any pages that don’t fit into the cache any longer. This is because a single write may evict multiple pages. If we have space for 32KB, and we currently have 4 separate pages in the cache, we are full. Adding a write for 2 pages, however, will need to evict two items out of the cache. Another aspect that I changed from the last post is that I’m no longer requiring that the current page will not be evicted immediately.

I mentioned that I changed the Lru’s implementation significantly.  Previously it was fairly generic, but it turns out that it is better to specialize it specifically for our needs. Here is the current version:

This Lru is implemented using a generational approach. The idea is that we can specify the sizes of the young and old generations independently. A new page that is added to the cache is written to the young generation. If there is another write to the page, it will be moved to the old generation. If the size of the data in each generation is too big, we use a simple Least Recently Used to evict items. An item that expired on the old generation will be moved to the young generation, while an item that expires from the young generation will be evicted completely. Here is how this looks like in code:

This is the same rough algorithm used by MySQL. The advantage here is that there is O(1) cost for the entire process, but at the same time, busy pages actually have a good chance of not being written to the disk (which is ideal, since if they are busy, they keep changing). To give some context to the sizes involved, we may decide that the old generation contains 192 MB and the young generation is 64 MB. A page write will remain in the young generation until we have another 64 MB (8,192 individual page writes), if there were any writes to it again in that time frame, it will go to the old generation. Only if we didn't have any writes to it in that time frame will we flush it to disk. Once a page was moved to the old generation, it has to go through a buffer of 256 MB (32K page writes) before it will go to the disk. Hot pages should be safe from spurious writes, especially since we’ll write them to the disk anyway whenever we do a checkpoint.

When a page is evicted from the cache, the writePageToDisk() method is called on it, there is a bit of work going on here:

An important observation about writing to the disk is that coalescing writes can have a huge impact on performance. Instead of having to go to the disk multiple times, we can schlepp the data there just once. The amount of the data we write is actually far less significant than what you may expect. You can see that the first thing that we do is to try to merge the pages (both forward and backward). If we find a candidate, we expand our own WriteState and release the one that subsumed. This is an important aspect, because the writes we send to the disk don’t necessarily match the old / young generations division. If we evicted a page that is adjacent to another modified page, we’ll write them both to the disk. That is even if the second page is in the old generation (and that will evict the page from the write cache). In practice, databases exhibit a lot of locality, so I don’t expect that to be an actual issue.

The last thing we do before actually submitting the I/O operation is to register the write in the inflight record. This is basically a sorted list (by the first page for that write) which we’ll check on any future writes. We also increment the number of background writes we have so we’ll know to wait for them to complete. Note that if we fail to submit the I/O, we decrement the number of backgroundWrites, we use this value to wait for async operations (for example, during shutdown). And since such errors are probably fatal, we are heading that way soon.

The actual I/O happens in the background and we’ll call the callback function when that is done (either successfully or with an error). So far, all the code we saw was single threaded and protected by the lock mutex. The callback, on the other hand, is running without any such protections. Furthermore, the callback thread is running all the notifications we have. We want to do as little as possible there. Here is what I ended up with:

We register the completed write in the completed linked list and wake any pending waiters. Any actual processing, of course, is moved to the part of the process where we actually run under lock and don’t have to worry about concurrency and multi threading. The actual behavior around completing writes is implemented in the drainCompleted() call that is part of the write() call.

Even though we are under lock, the I/O completion may compete with us on the completed list, so we’re using lock free methods to iterate through the list of completed actions. The actual behavior is straightforward, we remove the write from the list of inflight writes and return an error if the write had a problem. Note that we have to deal with potentially nested errors as well.  Aside from calling this as part of the write call(), we may also want to explicitly wait for all the pending operations to complete. This is done using the waitForWrites() call:

We drain all the completed writes and then wait for any pending ones to complete as well. We repeat this until there are no more backgroundWrites. Note that we run this under the lock, so we know that there can never be additional writes happening.

In special circumstances, we may want to ask the writer to push everything to the disk, we do that by calling the checkpoint() method:

There is almost nothing there, I’m happy to say, we are almost done. We evict pages from the cache from the youngest to the oldest, using the same write coalescing as before. Given write locality, that is likely to produce good results. We complete the process by waiting for all those writes to complete, and we are done.

There is just one last topic that I want to cover: Shutting  down:

Shutting down turns out to be a fairly involved process. This is especially the case if we are shutting down because of an error. We need to wait until all the pending writes have been completed (to do otherwise is to invite Use-After-Free bugs). That is why we call the waitForWrites() until it completes successfully. At worst, each time this is called will process a single write.

On shutdown, we can’t really start writing to the disk (after all, we may be shutting down because of the disk). We just discard the data in this case. That is something that we are okay to do, because we assume that any important data can be restored when we run recovery. This is one of those cases where the fact that we are building a database makes our life easier. We don’t need to try very hard to persist to disk, we can assume that as a given.

I’m pretty happy with how all this turned out. The full code for the pager is now at around 1,500 lines of code. I think that at this point we are pretty much at or near what you’ll need to replace the mmap() based pager that I demonstrated earlier in this series. This is the sort of code that needs a metric ton of tests, of course, and a lot of actual experience throwing stuff at it. The nice thing about this approach is that this is entirely in our hands. We can now implement whatever policies we want. The actual behavior is more or less done, but we can play with the exact mechanics and what will trigger what at will. In some cases, that can give us a great deal of power and flexibility.

I read this post and it took me very little time to spot a pretty nasty bug. Here is the relevant section:

The bug is on line #4. The code assumes that a call to read() will return less than the requested number of bytes only at the end of the file. The problem with that approach is that this is explicitly documented to not work this way:

It is not an error if the returned value n is smaller than the buffer size, even when the reader is not at the end of the stream yet. This may happen for example because fewer bytes are actually available right now (e. g. being close to end-of-file) or because read() was interrupted by a signal.

This is a super common error in many cases. And in the vast majority of the cases, that would work. Except when it wouldn’t.

The underlying implementation of File::read() will call read() or ReadFile(). ReadFile() (Windows) is documented to read as much as you requested, unless you hit the end of file. The read() call, on Unix, is documented to allow returning less than requested:

It is not an error if this number is smaller than the number of bytes requested

Aside from signals, the file system is free to do a partial read if it has some of the data in memory and some not. I’m not sure if this is implemented in this manner, but it is allowed to do so. And the results for the code above in this case are absolutely catastrophic (decryption will fail, encryption will emit partial information with no error, etc).

I’m writing this blog post because reading the code made the error jump at me. Was bitten by this assumption too many times.

In the previous post I outlined some ideas about how to implement a more efficient write behind. The idea is that whenever we write pages to the Pager, we’ll not trigger an immediate write to the disk. Instead, we’ll keep the data in memory and only write to the disk when we hit a certain threshold. In many write scenarios, there are certain pages that are modified a lot (like the root page in a B+Tree) and pages that are modified rarely (a leaf page that got entries and will not be modified again). There is no point in writing the popular page to the disk, we’ll likely get another write to them shortly anyway. That calls to a Least Frequently Modified approach.

We don’t need to use a more complex approach, (like the Clock Sweep algorithm we use for the Pager), because we don’t have to deal with the same scenarios. There are not likely to be cases similar to scans, which throws a lot of complexities of buffer pool implementations. Writes operations are far more predictable in general and follow a pretty strict power law distribution. The task is simple: we have the list of pages that were modified, and at capacity, we’ll select some to send to the disk. The question is how to make that decision.

The simplest option is to go with the least recently used model. That is trivial to implement, The idea is that we have the following sequence of writes (assuming we have a capacity of 4 pages):

1, 2, 3, 2, 3, 1, 2, 3, 4, 2, 1, 2, 3, 4, 4, 2, 1, 6

In this case, the page that will be selected for eviction is #3, since it wasn’t modified the longest. The other alternative is to use least frequently used, in which case we have the following frequency table:

In this case, we’ll want to select page #4 for eviction. Since it is the one least used. (We don't consider #6 because it is the one we just inserted). I can make arguments for both sides, to be frank. It makes sense that the least frequently used is going to be the most relevant, right? The problem is that we need to also account for decaying usage over time. What do I mean by this?

We may have a page that is very hot, it gets used a lot for a certain period of time. After that point, however, it is no longer being written to, but because it was frequently used, it will take a long time to evict from the pool. A good example of such a scenario is when we have a B+Tree and we are inserting values in ascending orders. All the values for the tree are going to be placed in the same page, so if we have a lot of inserts, that page is going to be hot. Once it is full, however, we’ll start using another page as the target and then the page will reside in memory until some other page will have more usage.  A good discussion of least frequency used implementation is in this blog post.

A nice way to deal with the issue of decaying priorities over time in an LFU setup is to use the following formula to compute the priority of the pages:

The idea is that we compute the priority of the page based on the last access, so we are very close to the most recently used option. However, note that we compute the distance between accesses to the page. A page that is infrequently accessed will have low usages and a high delta sum. That will reduce its priority. Conversely, a page that is heavily used will have a low delta sum and high usage, so its value will be near the top.

Another option is to go with another clock sweep option. In this case, we use a simple least recently used model, but we keep count of the frequency of usages. In that case, if we have to evict a page that is heavily used, we can reduce its usages and give it another chance. The advantage here is that this is a far simpler model to work with, but gives roughly the same results. Another option we have is to use the midpoint insertion LRU.

There is also another consideration to take. The I/O cost isn’t linear. If I’m writing page #3 to disk, it is basically free from my perspective to write nearby pages. It is the same exact cost, after all, so why not do that?

We’ll need to write our own doubly linked list. The Zig’s standard library only contains a single linked list. It doesn’t take long to write such a data structure, but it is fun to do so. I absolutely get why implementing linked lists used to be such a common practice in interviews:

There isn’t really much to discuss here, to be honest. There is a bunch of code here, but it is fairly simple. I just had to implement a few operations. The code itself is straightforward. It is a lot more interesting when we see it being used to implement the LRU:

The push() method is where it all happens. We have two options here:

• We have a page already inside the LRU. In that case, we increment its usage counter and move it to the front of the list.
• This is a new page, so we have to add it to the LRU. If there is enough capacity, we can just add it to the front of the list and be done with it.

However, things get interesting when we are at capacity. At that point, we actually need to select a page to evict. How can we do that? We scan the end of the list (the oldest page) and check its usage. If it has more than a single usage, we half its usage counter and move it to the front. We continue to work on the tail in this manner. In essence, high usage counter will get reset rather quickly, but this will still give us a fairly balanced approach, with more popular pages remaining in the pool for longer.

When we evict a page, we can return it back to the caller, which can then write it to the disk. Of course, you probably don’t want to just write a single page. We need to check if we have additional pages nearby, so we can consolidate all of them at once to the disk.

I’ll touch on that in my next post.

In the last blog post I presented the manner in which the Pager can write data to disk. Here is a reminder:

We acquire the writer (and a lock on it), call write on the pages we want to write on (passing the buffer to write on them), and finalize the process by calling flushWrites() to actually write the data to disk. As a reminder, we assume that the caller of the Pager is responsible for coordination. While we are writing to a specific page, it is the responsibility of the caller to ensure that there are no reads to that page.

The API above is intentionally simplistic , it doesn’t give us a lot of knobs to play with. But that is sufficient to do some fairly sophisticated things. One of the interesting observations is that we split the process of updating the data file into discrete steps. There is the part in which we are updating the in memory data, which allows other threads to immediately observe it (since they’ll read the new details from the Pager’s cache). Separately, there is the portion in which we write to the disk. The reason that I built the API in this manner is that it provides me with the flexibility to make decisions.

Here are some of the things that I can do with the current structure:

• I can decide not to write the data to the disk. If the amount of modified pages is small (very common if I’m continuously modifying the same set of pages) I can skip the I/O costs entirely and do everything in memory.
• Flushing the data to disk can be done in an asynchronous manner. In fact, it is already done in an asynchronous manner, but we are waiting for it to complete. That isn’t actually required.

The way the Pager works, we deposit the writes in the pager, and at some future point the Pager will persist them to disk. The durability aspect of a database is not reliant on the Pager, it is a property of the Write Ahead Log, usually.

If I wanted to implement a more sophisticated approach for writing to the disk, I could implement a least recently used cache for the written pages. When the number of pages in memory exceeds a certain size, we’ll start writing the oldest to disk. That keeps the most used pages in memory and avoids needless I/O.  At certain points, we can ask the Pager to flush everything to the disk, this gives us a checkpoint, where we can safely trim the Write Ahead Log. A good place to do that is whenever we reach the file size limit of the log and need to create a new one.

So far, by the way, you’ll notice that I’m not actually talking about durability, just writing to the disk. The durability aspect is coming from something we did long ago, but didn’t really pay attention to. Let’s look at how we are opening files, shall we:

Take a look at the flags that we pass to the open() command, we are asking the OS to use direct I/O (bypassing the buffer pool, since we’ll use our own) as well as using DSYNC write mode. The two together means that the write will skip any buffering / caching along the way and hit the disk in a durable manner. The fact that we are using async I/O means that we need to ensure that the buffers we write are not modified while we are saving them. As we currently have the API, there is a strong boundary for consistency. We acquire the writer, write whatever pages we need and flush immediately.

A more complex system would be needed to manage higher performance levels. The issue is that in order to do that, we have to give up a level of control. Instead of knowing exactly where something will happen, we can have a more sophisticated approach, but  we’ll need to be aware that we don’t really know at which point the data will be persisted.

At this point, however, there is a good reason to ask, do we even need to write durably? If we are limiting the consistency of the data to specific times requested by the caller (such as when we replace the Write Ahead Log), we can just call fsync() at the appropriate times, no? That would allow us to use buffered writes from most I/O.

I don’t think that this would be a good idea. Remember that we are using multiple files. If we’ll use buffered I/O and fsync(), we’ll need to issue multiple fsync() calls, which can be quite expensive. It also means higher memory usage on the system because of the file system cache, for data we determine is no longer in memory. It is simpler to use direct I/O for the whole thing, after all.

In the next post, I’m going to show how to implement a more sophisticated write-behind algorithm and discuss some of the implications of such a design.

At long last, we are now at the point where we can write data back to the disk.  Before we can do that, however, we need to figure out what sort of writes we want to allow. The idea that I have in mind for the Pager is to follow the same path as Voron does. In other words, for writes, we can make the following assumptions:

• There is only a single thread doing writes to the pager.
• There are no readers for the writes until the write is completed.
• There is a distinction between writing the data to the pager and writing the data to the disk.

Let’s break those assumptions apart and see what they bring to the table. The fact that we can assume only a single writer thread at any given point is pretty important. It means that the level of complexity that we have to face is greatly reduced. In the same sense, the fact that we don’t need to deal with concurrent readers or any consistency boundary for the data while it is being written will greatly simplify things for us. Finally, we make a distinction between writing to the pager and writing to the disk. Writing to the disk is _slow_, so we want to avoid doing that at any critical areas and push that to the background.

Finally, there is another aspect to consider. Internally, the Pager works with 2MB chunks, but to the outside world, it is using 8KB pages. When we write, we always write at the 8KB pages, not chunks. How would that work for the Pager?

The Pager itself is concurrent, but we only allow a single writer at a time, we can achieve this by centralizing all the write activities in the Writer struct, like so:

For now, I’m going to ignore the fields in the Writer struct, we’ll touch on them in detail later. In order to use the writer, you need to acquire it, write as many pages as you need, then release it. Here is a usage example:

The basic idea is fairly simple. With the writer, we operate at the page boundary to write as many pages as we need, once we are done, the call to flushWrites() persists the data to disk and then we can release the writer. Let’s dig a bit deeper and see how that works, shall we?

The write() call is about as basic as you can get. We use the getPage() function to get the right page, memcpy the data and that is about it, right? There are only two other things here that are important:

• We record which chunk (the 2MB chunk of memory, out of which we carve the 8KB pages) at the writer’s level, is using the loadedChunksForWrites value.
• We remember which pages we wrote to using the writtenPages hash table.

This is intentionally bare bones, because that is actually sufficient for our needs. The  fact that we remember which chunks we loaded (and keep a reference to them) will prevent us from reclaiming them, so even though we just wrote to memory, another thread can get the data and start using it without waiting for the disk. Of course, we still need to hit the disk eventually, that is what flushWrites() is about.

There is a lot that is going on here, let’s break it up. We start by allocating a temporary array and copying the keys from the writtenPages hash table to it. We then sort the array. This is done so we’ll be able to process the writes in a sequential manner, which is likely to be faster, even with async I/O. We then scan the list of pages in order, trying to merge writes together. The idea is to issue the minimum number of write calls. Finally, we’ll wait for all the writes to complete. Okay, maybe it isn’t that complex.  There is a bunch of code here, but it is mostly straightforward. Note that we also prepare the writeResults list to accept the results of the write to the disk.

As for writing to the disk, this is done using the PagerRing we previously looked at:

To write a buffer to the disk, we simply get the buffer from the Pager (reusing all the work we did in getPage()), increment the number of outstanding writes and then submit the work for the ring for processing. We setup the completeFlush as the callback function on completion. The PagerRing will call us when it is done writing to the disk. If there is an error, we’ll record it and reduce the number of outstanding writes. If there are no more outstanding writes, we’ll wake any waiters. That part is handled in the waitForAllDiskWritesToComplete().

We start by waiting for the outstanding writes to complete, waiting if needed. Then we can reset the state of the Writer. We start by resetting the written pages and then iterate over all the loaded chunks and release them. After the call, the Pager may decide to remove them from memory. This is fine, since they were already written to disk.

Except… if there was an error. You might have noticed that we are gathering the errors on each individual write operation we send, but we are actually only looking at the first one. For that matter, we clear the state of the Writer regardless if there were errors or not.

In general, an I/O error from the disk is not something that is recoverable. What you can do at this stage is to raise the error higher and run whatever recovery you have on startup.

In the next post, I’m going to be talking about durability and the overall expected performance of the system under this sort of model.

I was pointed to this paper on twitter: Are You Sure You Want to Use MMAP in Your Database Management System?

As you can imagine, this is a topic near and dear to my heart. This is especially the case since I am currently writing the Implementing a file pager in Zig posts series. I implemented the same low level mechanics using mmap, using mmap, I have < 100 lines of code and can start building higher level concepts almost immediately. Writing my own pager is currently a 10 posts series and the end doesn’t seem to be in sight.

I’m going to use this post to respond to the article. As a reminder, I’m the founder of RavenDB and I wrote Voron, a mmap based storage engine, and has been running that across hundreds of clients and literally tens of millions of instances in production. I am also writing a book about building a storage engine that uses mmap internally.

The paper itself does a great job of outlining the issue of using mmap as the buffer pool in DBMS. What it doesn’t cover, however, is the alternative. I will touch on specific points from the paper shortly, but I want to point out that the article compares apples to camels in the benchmarks and conclusions. Note that I don’t necessarily disagree with some of the statements, mmap certainly has challenges that you need to deal with, but if you avoid that, you can’t have wave everything that it brings to the table.

When building a database, using mmap has the following advantages, the OS will take care of:

• Reading the data from disk
• Concurrency between different threads reading the same data
• Caching and buffer management
• Eviction of pages from memory
• Playing nice with other processes in the machine
• Tracking dirty pages and writing to disk*

I put an asterisk on the last one because it probably requires your attention as well.

If you aren’t using mmap, on the other hand, you still need to handle all those issues. That is a key point that I believe isn’t addressed in the paper. Solving those issues properly (and efficiently) is a seriously challenging task. Given that you are building a specialized solution, you can probably do better than the generic mmap, but it will absolutely have a cost. That cost is both in terms of runtime overhead as well as increased development time.

The comparison that was made by the paper was done using fio benchmark tool, which is great if you want to test your storage system, but is pretty much irrelevant if you are trying to benchmark a buffer pool. Consider the following:

For the mmap version, we need to compute the address of the page and that is pretty much it. For the manual buffer pool, the list of tasks that we need to handle is long. And some of them require us to be thread safe. For example, if we handed a page to a transaction, we need to keep track of that page status as being in use. We cannot evict this page until the transaction is done with it. That means that we probably need to do atomic reference counting, which can have very high costs. There are other alternatives, of course, but they all have even higher degrees of complexity.

In practice, data access within a database isn’t actually random, even if you are doing random reads. There are pages that are going to almost always be referenced. The root page in the B+Tree is a good example. It is always going to be used. Under atomic reference counting, that page is going to be a bottleneck.

Ignoring such overhead of the buffer pool management means that you aren’t actually comparing equivalent structures. I should also point out that I’m probably forgetting a few other tasks that the buffer pool needs to manage as well, which complicate its life significantly. Here is an example of such a buffer pool implementation from what is effectively a random GitHub repository. You can see what the code is trying to do here. The reason I point to this is that there is a mutex there (and I/O under the lock), which is fairly typical for many buffer pools. And not accounting for the overhead of buffer pool management is seriously skewing the results of the paper.

All of this said, I absolutely agree that mmap can be challenging. The paper outlines 4 different problems, which I want to address.

Problem #1 – Transactional safety

A database needs to know when the data is persisted to disk. When using mmap, we explicitly give up that knowledge. That can be a challenge, but I don’t see that as a seriously different one from not using mmap. Let’s consider the manner in which Postgres is working. It has its own buffer pool, and may modify the pages as a result of a write. Postgres may need to evict modified pages to disk before the transaction that modified them is committed. The overhead of managing that is just… part of the challenge that we need to deal with.

For RavenDB, as the paper points out, we modify the pages outside of the mmap memory. This is actually not done for the reason the paper describes. I don’t actually care if the data is written to memory behind my back. What I care about is MVCC (a totally separate concern than buffer management). The fact that I’m copying the modified data to the side means that I Can support concurrent transactions with far greater ease. In a similar fashion, Postgres handles MVCC using multiple entries for the same row in the same page.

When the transaction commits and older transactions no longer need the old version of the data, I can push the data from the modified buffers to the mmap region. That tends to be fairly fast (given that I’m basically doing memcpy(), which runs at memory speed) unless I have to page data in, more on that later.

The paper mentions the issue of single writer in LMDB, and I wanted to point out that a single writer model is actually far more common (and again, not really related to the buffer pool issue). Off the top of my head, most embedded databases implement a single writer model.

• LMDB
• Voron (RavenDB’s storage engine)
• LevelDB
• Lucene

The one that I can think that doesn’t have a single writer is RocksDB(where allow_concurrent_memtable_write is for writes to the memtable, not related to file I/O).

The reason this matters is that embedded systems can typically assume that all operations in a transaction will complete as a unit. Compare to Postgres, where we may have a transaction spanning multiple network calls, interleaving writes is a must. If we could avoid such concurrency, that would be far preferable. You can get additional concurrency by having sharding writes, but that is usually not needed.

Problem #2 – I/O Stalls

The paper points out, quite correctly, that not having control over the I/O means that you may incur a page fault at any time. In particular, you may end up blocked on I/O without really noticing. This can be a killer especially if you are currently holding a lock and blocked on page fault. Indeed, I consider this to be the most serious issue that you have to deal with mmap based systems.

In practice, however, the situation isn’t so clear cut. Until quite recently, the state of asynchronous I/O on Linux was quite iffy. Until the arrival of io_uring, certain operations that you expected to be async would block occasionally, ruining your day. The paper mentions that you can use async I/O to issue I/O requests to load the next pages (non sequentially) from the disk when you are performing certain operations. You can do the same with mmap as well, and RavenDB does just that. When you start a scan on a B+Tree, RavenDB will ask the OS to ensure that the memory we are interested in is in memory before we actually get to it. On Linux, this is done with madvise(WILL_NEED) call. That call may be blocking, so we actually have a dedicated thread that is meant to handle such a scenario.  In practice, this isn’t really that different from how you’ll handle it with async I/O.

Another consideration to deal with is the cost of mapping at the kernel level. I’m not talking about the I/O cost, but if you have many threads that are faulting pages, you’ll run into problems with the page table lock. We have run into that before, this is considered an OS level bug, but it obviously has an impact on the database. In practice, however, the overhead of memory management is the same in most cases. If you are reading via mmap or allocating directly, you’ll need to orchestrate things. Note that the same page table lock is also in effect if you are heavily allocating / freeing, since you’re also modifying the process page table.

Problem #3 – Error Handling

Error handling is a serious concern for a database. The paper points out that databases such as SQL Server may run a checksum when reading data from disk. When you use a buffer pool, the boundary of reading from the disk is obvious and you can easily validate the read from the disk. Voron is using mmap exclusively, and we do have checksums. We validate the page from the disk the first time that we access it (there is an internal bitmap that is used for that).  There isn’t a big difference between the two anyway. We only check a given page once per run, because to do otherwise is meaningless. When you use read() to get data from the disk, you have no guarantees that the data wasn’t fetched from a cache along the way. So you may validate the data you read is “correct”, while the on disk representation is broken. For that reason, we only do the check once, instead of each time.

A far greater issue to deal with is I/O errors. What do you do when a read or a write fails? If you are using system calls to manage that, you get a return code and can react accordingly. If you are using mmap, the system will generate a SIGBUS that you’ll have to (somehow) handle.

For a database, dealing with I/O errors has a single correct answer. Crash and then run recovery from scratch. If the I/O system has returned an error, there is no longer any way to know what the state of that is. See: fsync-gate. The only way to recover is to stop, reload everything (apply the WAL, run recovery, etc) and get back into a stable state. SIGBUS isn’t my cup of tea with regards to handling this, but error handling for I/O error isn’t actually something that you do, so just restarting the process ends up more acceptable than you might initially think.

Problem #4 – Performance issues

The paper points out three common reasons for performance issues with mmap usage:

1. page table contention
2. single threaded page eviction
3. TLB shootdowns

The first issue is something that I have run into in the past. It was a bug in the operating system which was fixed. There is no longer a single page table in both Windows and Linux.

The single threaded eviction, on the other hand, is something that we never run into. When using Voron, we map the memory using MAP_SHARED, and most of the time, the memory isn’t dirty. If the system needs memory, it can do that when it assigns a page by just discarding the memory of an unmodified shared page. In this model, we typically see most of the memory as shared, clean. So there isn’t a lot of pressure to evict things, and it can be done on as needed basis.

The TLB shootdown issue is not something that we ever run into as a problem. We have run TB range databases on Raspberry PI with 4GB of RAM and hammered that in benchmarks (far exceeding the memory capacity). The interesting thing here is that the B+Tree nature means that the upper tiers of the tree were already in memory, so we mostly ended up with a single page fault per request. In order to actually observe the cost of TLS Shootdown in a significant manner, you need to have:

• really fast I/O
• working set that significantly exceeds memory
• no other work that needs to be done for processing a request

In practice, if you have really fast I/O, you spent money on that, you’ll more likely get more RAM as well. And you typically need to do something with the data you read, which means that you won’t notice the TLB shootdown as much.

Finally, going back to how I started this post. This assumes that there are no other costs of not using mmap and using direct IO. The benchmark doesn’t account for those extra costs. For example, without mmap, who is doing evictions? In practice, that will lead to the same sort of considerations that you’ll have when dealing with mmap. This is especially the case with TLS shootdown when we start talking about high memory traffic (which likely modifies page allocations for the process, leading to the same scenario).

The paper has been quite interesting to read and it has presented a number of real problems that occur with mmap based systems, but I’m afraid that it doesn’t present the alternatives properly and vastly underestimates both costs and complexity of not using mmap and writing your own buffer pool.

Kamran Ayub has another great article discussing how to utilize RavenDB to analyze crypto markets data.

In addition to the article, you can see the full source code for the sample application. There is even a really cool intro video for the application and article:

Up to this point, we focused on reading data from the disk, we can do that up to a point. Eventually we’ll run out of memory (assuming that the database is bigger than memory, which is a pretty safe assumption). That means that we need to decide what to remove from memory. When we use mmap(), the OS gets to decide that for us. In fact, that is probably the best argument against using mmap(). The additional control we get from knowing how to manage the memory is the chief reason to take the plunge and manage our own memory.

There are a lot of algorithms around managing memory, I really like this one, because it is quite elegant. However, that requires quite a lot of states to be dealt with, especially when working with highly concurrent systems. Instead, I chose to look at the clock sweep algorithm. This is also implemented by PostgreSQL, and it is actually far simpler to work with. The idea is that for each page, we maintain a usage count. Each time that we need to get a page, we’ll increment its usage count (up to a small limit). Each time we need to evict a page, we’ll search for a page that can be evicted and has no recent usage. If it has usages, we’ll decrement that value and repeat until we find something.

Our buffer management isn’t actually dealing with pages, however. We are working with 2MB chunks, instead. The principal is the same, but using bigger aggregates is advantageous given typical memory sizes these days. The first thing that we need to do is to modify the ChunkMetadata. I’m showing only the relevant changes.

The major change here is the introduction of the usages field. That is a 3 bits field (0 .. 8 in range) and we reduced the number of references a chunk can have to about 500 million (should be sufficient, I believe). The idea here is that each time that we call addRef(), we’ll increment the usages count, like so:

Zig has a nice feature that I’m using here, saturating addition. In other words, if the value is incremented beyond its limit, it is clamped to the limit. That means that I don’t have to worry about overflows, etc. I took a look at how this is implemented, and the compiler generates the following assembly for this code (x +| 100) :

This may look daunting, but I’ll break it down. First, we add the 100 to the value, as you can expect (the value is currently in the EAX register). Then we store –1 (value of 0xFFFFFFFF) in the ECX register. Finally, we use the CMOV instruction (the CMOVB in the snipper is a variant of CMOV), telling it to store ECX in EAX if the carry flag is set on the last addition instruction. For fun, this also avoids a branch in the code, which is great for performance.

One of the critical functions that we need to consider here is the behavior of the pager when we are accessing rare pages once, and then never again. A really common scenario is when we are doing a scan of some data for a rare query. For that reason, the usages behavior is a bit more complex than one might imagine. Let’s explore this for a bit before moving on. Look at the following code, I marked the important lines with stars:

When we mark a chunk as loaded, we copy the usages from the current record. That starts out as zero, so for the scenario of accessing rare pages, we’ll have a good reason to evict them soon. However, there is a twist, when we remove a chunk from memory, we also set its usage count to 1. That is an interesting issue. The chunk is not loaded, why does it have a usage count? That is because if we removed it from memory, and we load it again, we want it to start with a higher usage count (and less chance to be evicted). In this manner, we are somewhat simulating the 2Q algorithm.

Now, let’s take a look at the actual reclaiming portion, shall we? In the chunk metadata, we have the following behavior:

If we have a value and no outstanding references, we can reclaim it, if we don’t have a value, we’ll reduce the usage count anyway. The idea is that when we unload a page, we’ll set the usages count to 1. The usage counter of an unloaded page will reset this as time goes by, so if there has been a sweep on an empty chunk that has a usage count, we’ll not count that for the next time we load it. Basically, after unloading a chunk, if we reload it soonish, we’ll keep it around longer the next time. But if it is only rarely loaded, we don’t care and will forget that it was loaded previously.

The process for actually reclaiming a chunk is shown here:

We look at a particular index in the chunks, and check if we can claim that. Following out previous behavior, there are a bunch of options here. We can either have an empty chunk that remembers the previous usage, which we can reduce, or we can actually try to reclaim the chunk. Of course, that isn’t that simple, because even if we found a candidate for reclaiming, we still need to reduce its usages count. Only if it has no usages will we be able to actually start the removal process. Finally, we have the actual scanning of the chunks in the file, shown here:

We scan the file from the provided start index and try to reclaim each chunk in turn. That may reduce the usages count, as we previously discussed. The actual process will continue until we have found a chunk to reclaim. Note that we are always claiming just a single chunk and return. This is because the process is repeatable. We start from a given index and we return the last index that we scanned. If the caller needs to free more than a single chunk, they can call us again, passing the last index that we scanned.

That is why this is called the clock sweep algorithm. We are sweeping through the chunks that we have in the system, reaping them as needed. The code so far is all in the same FileChunks instance, but the Pager actually deals with multiple files. How would that work? We start by adding some configuration options to the pager, telling us how much memory we are allowed to use:

We have both soft and hard limits here, because we want to give the users the ability to say “don’t use too much memory, unless you really have to”.  The problem is that otherwise, users get nervous when they see 99% memory being used and want to keep some free. The point of soft and hard limits is that this gives us more flexibility, rather than setting a lower than needed limit and getting memory errors with GBs of RAM to spare.

In the Pager, we have the loadChunksToTransaction() that we looked at in the previous post. That is where we read the chunk from the file. We are going to modify this method so will reserve the memory budget before we actually allocate it, like so:

As you can see, we reserve the memory budget, then actually allocate the memory (inside markLoading()). If there is a failure, we release the budget allocation and report the error. To manage the memory budget, we need to add a few fields to the Pager:

You can see that releaseChunkMemoryBudget() is pretty trivial. Simply release the memory budget and move on. Things are a lot more complex when we need to reserve memory budget, however. Before we dive into this, I want to talk a bit about the filesRelcaimSweeps field. That is an interesting one. This is where we’ll keep the last position that we scanned in the Pager (across all pages). However, why is that an array?

The answer is simple. The Pager struct is meant to be used from multiple threads at the same time. Under memory pressure, we are likely to need to evict multiple chunks at once. In order to avoid multiple threads scanning the same range of the Pager to find chunks to remove, I decided that we’ll instead have several sweeps at the same time. On startup, we’ll initialize them to a random initial value, like so:

In this manner, under load, each thread is likely to scan an independent portion of the Pager’s memory, which should avoiding competing on the same memory to evict. And with that behind us, let’s see how we can actually use this to evict memory:

There is a lot of code here, and quite a few comments, because this is choke-full of behavior. Let’s dissect this in detail:

We start by checking the memory budget and see if we are below the soft memory limits. We are doing this optimistically, but if we fail, we reset the state and then start to scan the Pager for chunks we can release. We do that by accessing one of the filesRelcaimSweeps values using the current thread id. In this way, different threads are likely to use different values and move them independently.

We find the relevant file for the index and start scanning for chunks to release. We’ll stop the process when one of the following will happen:

• We released a chunk, in which case we are successful and can return with glory.
• We didn’t find a chunk, but found candidates (whose usage count is too high to discard).

In the case of the second option, we’ll look for better options before getting back to them and discarding them.

If we are completely unable to find anything to release, we’ll check if we exceed the hard memory limit and error, or just accept the soft limit as, well… soft limit and allocate the budget anyway.

This happens as part of the process for loading chunks in the pager, so we’ll only need to release a single chunk at a time. For that reason, we remember the state so the next operation will start from where we left off. You can think about this as a giant set of hands that are scanning the range of chunks in memory as needed.

There are actually a few things that we can implement that would make this faster. For example, we always scan through all the chunks in a file. We could try to maintain some data structure that will tell us which pages have usage count to consider, but that is actually complex (remember, we are concurrent). There is also another factor to consider. The ChunkMetdata is 64 bits in size, and a FilesChunk struct contains an array of 4096 such values, totaling 32KB in size. It is actually cheaper to scan through the entire array and do the relevant computation on each candidate than try to be smart about it.

I think that this is it for now, this post has certainly gone for quite a while. In the next post in the series, I want to tackle writes. So far we only looked at reads, but I think we have all the relevant infrastructure at hand already, so this should be simpler.