RavenDB 7.0 introduces vector search using the Hierarchical Navigable Small World (HNSW) algorithm, a graph-based approach for Approximate Nearest Neighbor search. HNSW enables efficient querying of large vector datasets but requires random access to the entire graph, making it memory-intensive.

HNSW's random read and insert operations demand in-memory processing. For example, inserting 100 new items into a graph of 1 million vectors scatters them randomly, with no efficient way to sort or optimize access patterns. That is in dramatic contrast to the usual algorithms we can employ (B+Tree, for example), which can greatly benefit from such optimizations.

Let’s assume that we deal with vectors of 768 dimensions, or 3KB each. If we have 30 million vectors, just the vector storage is going to take 90GB of memory. Inserts into the graph require us to do effective random reads (with no good way to tell what those would be in advance). Without sufficient memory, performance degrades significantly due to disk swapping.

The rule of thumb for HNSW is that you want at least 30% more memory than the vectors would take. In the case of 90GB of vectors, the minimum required memory is going to be 128GB.

Cost reduction

You can also use quantization (shrinking the size of the vector with some loss of accuracy). Going to binary quantization for the same dataset requires just 3GB of space to store the vectors. But there is a loss of accuracy (may be around 20% loss).

We tested RavenDB’s HNSW implementation on a 32 GB machine with a 120 GB vector index (and no quantization), simulating a scenario with four times more data than available memory.

This is an utterly invalid scenario, mind you. For that amount of data, you are expected to build the HNSW index on a machine with 192GB. But we wanted to see how far we could stress the system and what kind of results we would get here.

The initial implementation stalled due to excessive memory use and disk swapping, rendering it impractical. Basically, it ended up doing a page fault on each and every operation, and that stalled forward progress completely.

Optimization Approach

Optimizing for such an extreme scenario seemed futile, this is an invalid scenario almost by definition. But the idea is that improving this out-of-line scenario will also end up improving the performance for more reasonable setups.

When we analyzed the costs of HNSW in a severely memory-constrained environment, we found that there are two primary costs.

1. Distance computations: Comparing vectors (e.g., cosine similarity) is computationally expensive.
2. Random vector access: Loading vectors from disk in a random pattern is slow when memory is insufficient.

Distance computation is doing math on two 3KB vectors, and on a large graph (tens of millions), you’ll typically need to run between 500 - 1,500 distance comparisons. To give some context, adding an item to a B+Tree of the same size will have fewer than twenty comparisons (and highly localized ones, at that).

That means reading about 2MB of data per insert on average. Even if everything is in memory, you are going to be paying a significant cost here in CPU cycles. If the data does not reside in memory, you have to fetch it (and it isn’t as neat as having a single 2MB range to read, it is scattered all over the place, and you need to traverse the graph in order to find what you need to read).

To address this issue, we completely restructured how we go about inserting nodes into the graph. We avoid serial execution and instead spawn multiple insert processes at the same time. Interestingly enough, we are single-threaded in this regard. We extract from the process the parts where it does the distance computation and loads the vectors.

Each process will run the algorithm until it reaches the stage where it needs to run distance computation on some vectors. In that case, it will yield to another process and let it run. We keep doing this until we have no more runnable processes to execute.

We can then scan the list of nodes that we need to process (run distance computation on all their edges), and we can then:

• Gather all vectors needed for graph traversal.
• Preload them from disk efficiently.
• Run distance computations across multiple threads simultaneously.

The idea here is that we save time by preloading data efficiently. Once the data is loaded, we throw all the distance computations per node to the thread pool. As soon as any of those distance computations are done, we resume the stalled process for it.

The idea is that at any given point in time, we have processes moving forward or we are preloading from the disk, while at the same time we have background threads running to compute distances.

This allows us to make maximum use of the hardware. Here is what this looks like midway through the process:

As you can see, we still have to go to the disk a lot (no way around that with a working set of 120GB on a 32GB machine), but we are able to make use of that efficiently. We always have forward progress instead of just waiting.

Don’ttry this on the cloud - most cloud instances have a limited amount of IOPS to use, and this approach will burn through any amount you have quickly. We are talking about roughly 15K - 20K IOPS on a sustained basis. This is meant for testing in adverse conditions, on hardware that is utterly unsuitable for the task at hand. A machine with the proper amount of memory will not have to deal with this issue.

While still slow on a 32 GB machine due to disk reliance, this approach completed indexing 120 GB of vectors in about 38 hours (average rate of 255 vectors/sec). To compare, when we ran the same thing on a 64GB machine, we were able to complete the indexing process in just over 14 hours (average rate of 694 vectors/sec).

Accuracy of the results

When inserting many nodes in parallel to the graph, we risk a loss of accuracy. When we insert them one at a time, nodes that are close together will naturally be linked to one another. But when running them in parallel, we may “miss” those relations because the two nodes aren’t yet discoverable.

To mitigate that scenario, we preemptively link all the in-flight nodes to each other, then run the rest of the algorithm. If they are not close to one another, these edges will be pruned. It turns out that this behavior actually increased the overall accuracy (by about 1%) compared to the previous behavior. This is likely because items that are added at the same time are naturally linked to one another.

Results

On smaller datasets (“merely” hundreds of thousands of vectors) that fit in memory, the optimizations reduced indexing time by 44%! That is mostly because we now operate in parallel and have more optimized I/O behavior.

We are quite a bit faster, we are (slightly) more accurate, and we also have reasonable behavior when we exceed the system limits.

        What about binary quantization?

I mentioned that binary quantization can massively reduce the amount of space required for vector search. We also ran tests with the same dataset using binary quantization. The total size of all the vectors was around 3GB, and the total index size (including all the graph nodes, etc.) was around 18 GB. That took under 5 hours to index on the same 32GB machine.

If you care to know what it is that we are doing, take a look at the Hnsw.Parallel file inside the repository. The implementation is quite involved, but I’m really proud of how it turned out in the end.

Let’s say I want to count the number of reachable nodes for each node in the graph. I can do that using the following code:

void DFS(Node start, HashSet<Node> visited) 
{
    if (start == null || visited.Contains(start)) return;
    visited.Add(start);
    foreach (var neighbor in start.Neighbors) 
    {
        DFS(neighbor, visited);
    }
}


void MarkReachableCount(Graph g)
{
   foreach(var node in g.Nodes)
   {
       HashSet<Node> visited = [];
       DFS(node, visisted);
       node.ReachableGraph = visited.Count;
   }
}

A major performance cost for this sort of operation is the allocation cost. We allocate a separate hash set for each node in the graph, and then allocate whatever backing store is needed for it. If you have a big graph with many connections, that is expensive.

A simple fix for that would be to use:

void MarkReachableCount(Graph g)
{
   HashSet<Node> visited = [];
   foreach(var node in g.Nodes)
   {
       visited.Clear();
       DFS(node, visisted);
       node.ReachableGraph = visited.Count;
   }
}

This means that we have almost no allocations for the entire operation, yay!

This function also performs significantly worse than the previous one, even though it barely allocates. The reason for that? The call to Clear() is expensive. Take a look at the implementation - this method needs to zero out two arrays, and it will end up being as large as the node with the most reachable nodes. Let’s say we have a node that can access 10,000 nodes. That means that for each node, we’ll have to clear an array of about 14,000 items, as well as another array that is as big as the number of nodes we just visited.

No surprise that the allocating version was actually cheaper. We use the visited set for a short while, then discard it and get a new one. That means no expensive Clear() calls.

The question is, can we do better? Before I answer that, let’s try to go a bit deeper in this analysis. Some of the main costs in HashSet<Node> are the calls to GetHashCode() and Equals(). For that matter, let’s look at the cost of the Neighbors array on the Node.

Take a look at the following options:

public record Node1(List<Node> Neighbors);
public record Node2(List<int> NeighborIndexes);

Let’s assume each node has about 10 - 20 neighbors. What is the cost in memory for each option? Node1 uses references (pointers), and will take 256 bytes just for the Neighbors backing array (32-capacity array x 8 bytes). However, the Node2 version uses half of that memory.

This is an example of data-oriented design, and saving 50% of our memory costs is quite nice. HashSet<int> is also going to benefit quite nicely from JIT optimizations (no need to call GetHashCode(), etc. - everything is inlined).

We still have the problem of allocations vs. Clear(), though. Can we win?

Now that we have re-framed the problem using int indexes, there is a very obvious optimization opportunity: use a bit map (such as BitsArray). We know upfront how many items we have, right? So we can allocate a single array and set the corresponding bit to mark that a node (by its index) is visited.

That dramatically reduces the costs of tracking whether we visited a node or not, but it does not address the costs of clearing the bitmap.

Here is how you can handle this scenario cheaply:

public class Bitmap
{
    private ulong[] _data;
    private ushort[] _versions;
    private int _version;


    public Bitmap(int size)
    {
        _data = new ulong[(size + 63) / 64];
        _versions = new ushort[_data.Length];
    }


    public void Clear()
    {
        if(_version++ < ushort.MaxValue)
            return;
            
        Array.Clear(_data);
        Array.Clear(_versions);
    }


    public bool Add(int index)
    {
        int arrayIndex = index >> 6;
        if(_versions[arrayIndex] != _version)
        {
            _versions[arrayIndex] = _version;
            _data[arrayIndex] = 0;
        }


        int bitIndex = index & 63;
        ulong mask = 1UL << bitIndex;
        ulong old = _data[arrayIndex];
        _data[arrayIndex] |= mask;
        return (old & mask) == 0;
    }
}

The idea is pretty simple, in addition to the bitmap - we also have another array that marks the version of each 64-bit range. To clear the array, we increment the version. That would mean that when adding to the bitmap, we reset the underlying array element if it doesn’t match the current version. Once every 64K items, we’ll need to pay the cost of actually resetting the backing stores, but that ends up being very cheap overall (and worth the space savings to handle the overflow).

The code is tight, requires no allocations, and performs very quickly.

We recently tackled performance improvements for vector search in RavenDB. The core challenge was identifying performance bottlenecks. Details of specific changes are covered in a separate post. The post is already written, but will be published next week, here is the direct link to that.

In this post, I don’t want to talk about the actual changes we made, but the approach we took to figure out where the cost is.

Take a look at the following flame graph, showing where our code is spending the most time.

As you can see, almost the entire time is spent computing cosine similarity. That would be the best target for optimization, right?

I spent a bunch of time writing increasingly complicated ways to optimize the cosine similarity function. And it worked, I was able to reduce the cost by about 1.5%!

That is something that we would generally celebrate, but it was far from where we wanted to go. The problem was elsewhere, but we couldn’t see it in the profiler output because the cost was spread around too much.

Our first task was to restructure the code so we could actually see where the costs were. For instance, loading the vectors from disk was embedded within the algorithm. By extracting and isolating this process, we could accurately profile and measure its performance impact.

This restructuring also eliminated the "death by a thousand cuts" issue, making hotspots evident in profiling results. With clear targets identified, we can now focus optimization efforts effectively.

That major refactoring had two primary goals. The first was to actually extract the costs into highly visible locations, the second had to do with how you address them. Here is a small example that scans a social graph for friends, assuming the data is in a file.

def read_user_friends(file, user_id: int) -> List[int]:
    """Read friends for a single user ID starting at indexed offset."""
    pass # redacted


def social_graph(user_id: int, max_depth: int) -> Set[int]:
    if max_depth < 1:
        return set()
    
    all_friends = set() 
    visited = {user_id} 
    work_list = deque([(user_id, max_depth)])  
    
    with open("relations.dat", "rb") as file:
        while work_list:
            curr_id, depth = work_list.popleft()
            if depth <= 0:
                continue
                
            for friend_id in read_user_friends(file, curr_id):
                if friend_id not in visited:
                    all_friends.add(friend_id)
                    visited.add(friend_id)
                    work_list.append((friend_id, depth - 1))
    
    return all_friends

If you consider this code, you can likely see that there is an expensive part of it, reading from the file. But the way the code is structured, there isn’t really much that you can do about it. Let’s refactor the code a bit to expose the actual costs.

def social_graph(user_id: int, max_depth: int) -> Set[int]:
    if max_depth < 1:
        return set()
    
    all_friends = set() 
    visited = {user_id} 
    
    with open("relations.dat", "rb") as file:
        work_list =  read_user_friends(file, [user_id])
        while work_list and max_depth >= 0:
            to_scan = set()
            for friend_id in work_list: # gather all the items to read
                if friend_id in visited:
                    continue
                
                all_friends.add(friend_id)
                visited.add(friend_id)
                to_scan.add(curr_id)


            # read them all in one shot
            work_list = read_users_friends(file, to_scan)
            # reduce for next call
            max_depth = max_depth - 1    
    
    return all_friends

Now, instead of scattering the reads whenever we process an item, we gather them all and then send a list of items to read all at once. The costs are far clearer in this model, and more importantly, we have a chance to actually do something about it.

Optimizing a lot of calls to read_user_friends(file, user_id) is really hard, but optimizing read_users_friends(file, users) is a far simpler task. Note that the actual costs didn’t change because of this refactoring, but the ability to actually make the change is now far easier.

Going back to the flame graph above, the actual cost profile differs dramatically as the size of the data rose, even if the profiler output remained the same. Refactoring the code allowed us to see where the costs were and address them effectively.

Here is the end result as a flame graph. You can clearly see the preload section that takes a significant portion of the time. The key here is that the change allowed us to address this cost directly and in an optimal manner.

The end result for our benchmark was:

• Before: 3 minutes, 6 seconds
• After: 2 minutes, 4 seconds

So almost exactly ⅓ of the cost was removed because of the different approach we took, which is quite nice.

This technique, refactoring the code to make the costs obvious, is a really powerful one. Mostly because it is likely the first step to take anyway in many performance optimizations (batching, concurrency, etc.).

Orleans is a distributed computing framework for .NET. It allows you to build distributed systems with ease, taking upon itself all the state management, persistence, distribution, and concurrency.

The core aspect in Orleans is the notion of a “grain” - a lightweight unit of computation & state. You can read more about it in Microsoft’s documentation, but I assume that if you are reading this post, you are already at least somewhat familiar with it.

We now support using RavenDB as the backing store for grain persistence, reminders, and clustering. You can read the official announcement about the release here, and the docs covering how to use RavenDB & Microsoft Orleans.

You can use RavenDB to persist and retrieve Orleans grain states, store Orleans timers and reminders, as well as manage Orleans cluster membership.

RavenDB is well suited for this task because of its asynchronous nature, schema-less design, and the ability to automatically adjust itself to different loads on the fly.

If you are using Orleans, or even just considering it, give it a spin with RavenDB. We would love your feedback.

RavenDB is moving at quite a pace, and there is actually more stuff happening than I can find the time to talk about. I usually talk about the big-ticket items, but today I wanted to discuss some of what we like to call Quality of Life features.

The sort of things that help smooth the entire process of using RavenDB - the difference between something that works and something polished. That is something I truly care about, so with a great sense of pride, let me walk you through some of the nicest things that you probably wouldn’t even notice that we are doing for you.

RavenDB Node.js Client - v7.0 released (with Vector Search)

We updated the RavenDB Node.js client to version 7.0, with the biggest item being explicit support for vector search queries from Node.js. You can now write queries like these:

const docs = session.query<Product>({collection: "Products"})
   .vectorSearch(x => x.withText("Name"),
      factory => factory.byText("italian food"))
  .all();

This is the famous example of using RavenDB’s vector search to find pizza and pasta in your product catalog, utilizing vector search and automatic data embeddings.

Converting automatic indexes to static indexes

RavenDB has auto indexes. Send a query, and if there is no existing index to run the query, the query optimizer will generate one for you. That works quite amazingly well, but sometimes you want to use this automatic index as the basis for a static (user-defined) index. Now you can do that directly from the RavenDB Studio, like so:

You can read the full details of the feature at the following link.

RavenDB Cloud - Incidents History & Operational Suggestions

We now expose the operational suggestions to you on the dashboard. The idea is that you can easily and proactively check the status of your instances and whether you need to take any action.

You can also see what happened to your system in the past, including things that RavenDB’s system automatically recovered from without you needing to lift a finger.

For example, take a look at this highly distressed system:

As usual, I would appreciate any feedback you have on the new features.

One of the more interesting developments in terms of kernel API surface is the IO Ring. On Linux, it is called IO Uring and Windows has copied it shortly afterward. The idea started as a way to batch multiple IO operations at once but has evolved into a generic mechanism to make system calls more cheaply. On Linux, a large portion of the kernel features is exposed as part of the IO Uring API, while Windows exposes a far less rich API (basically, just reading and writing).

The reason this matters is that you can use IO Ring to reduce the cost of making system calls, using both batching and asynchronous programming. As such, most new database engines have jumped on that sweet nectar of better performance results.

As part of the overall re-architecture of how Voron manages writes, we have done the same. I/O for Voron is typically composed of writes to the journals and to the data file, so that makes it a really good fit, sort of.

An ironic aspect of IO Uring is that despite it being an asynchronous mechanism, it is inherently single-threaded. There are good reasons for that, of course, but that means that if you want to use the IO Ring API in a multi-threaded environment, you need to take that into account.

A common way to handle that is to use an event-driven system, where all the actual calls are generated from a single “event loop” thread or similar. This is how the Node.js API works, and how .NET itself manages IO for sockets (there is a single thread that listens to socket events by default).

The whole point of IO Ring is that you can submit multiple operations for the kernel to run in as optimal a manner as possible. Here is one such case to consider, this is the part of the code where we write the modified pages to the data file:

using (fileHandle)
{
    for (int i = 0; i < pages.Length; i++)
    {
        int numberOfPages = pages[i].GetNumberOfPages();


        var size = numberOfPages * Constants.Storage.PageSize;
        var offset = pages[i].PageNumber * Constants.Storage.PageSize;
        var span = new Span<byte>(pages[i].Pointer, size);
        RandomAccess.Write(fileHandle, span, offset);


        written += numberOfPages * Constants.Storage.PageSize;
    }
}

PID     LWP TTY          TIME CMD
  22334   22345 pts/0    00:00:00 iou-wrk-22343
  22334   22346 pts/0    00:00:00 iou-wrk-22343
  22334   22347 pts/0    00:00:00 iou-wrk-22334
  22334   22348 pts/0    00:00:00 iou-wrk-22334
  22334   22349 pts/0    00:00:00 iou-wrk-22334
  22334   22350 pts/0    00:00:00 iou-wrk-22334
  22334   22351 pts/0    00:00:00 iou-wrk-22334
  22334   22352 pts/0    00:00:00 iou-wrk-22334
  22334   22353 pts/0    00:00:00 iou-wrk-22334
  22334   22354 pts/0    00:00:00 iou-wrk-22334
  22334   22355 pts/0    00:00:00 iou-wrk-22334
  22334   22356 pts/0    00:00:00 iou-wrk-22334
  22334   22357 pts/0    00:00:00 iou-wrk-22334
  22334   22358 pts/0    00:00:00 iou-wrk-22334

Actually, those aren’t threads in the normal sense. Those are kernel tasks, generated by the IO Ring at the kernel level directly. It turns out that internally, IO Ring may spawn worker threads to do the async work at the kernel level. When we had a separate IO Ring per file, each one of them had its own pool of threads to do the work.

The way it usually works is really interesting. The IO Ring will attempt to complete the operation in a synchronous manner. For example, if you are writing to a file and doing buffered writes, we can just copy the buffer to the page pool and move on, no actual I/O took place. So the IO Ring will run through that directly in a synchronous manner.

However, if your operation requires actual blocking, it will be sent to a worker queue to actually execute it in the background. This is one way that the IO Ring is able to complete many operations so much more efficiently than the alternatives.

In our scenario, we have a pretty simple setup, we want to write to the file, making fully buffered writes. At the very least, being able to push all the writes to the OS in one shot (versus many separate system calls) is going to reduce our overhead. More interesting, however, is that eventually, the OS will want to start writing to the disk, so if we write a lot of data, some of the requests will be blocked. At that point, the IO Ring will switch them to a worker thread and continue executing.

The problem we had was that when we had a separate IO Ring per data file and put a lot of load on the system, we started seeing contention between the worker threads across all the files. Basically, each ring had its own separate pool, so there was a lot of work for each pool but no sharing.

If the IO Ring is single-threaded, but many separate threads lead to wasted resources, what can we do? The answer is simple, we’ll use a single global IO Ring and manage the threading concerns directly.

Here is the setup code for that (I removed all error handling to make it clearer):

void *do_ring_work(void *arg)
{
  int rc;
  if (g_cfg.low_priority_io)
  {
    syscall(SYS_ioprio_set, IOPRIO_WHO_PROCESS, 0, 
        IOPRIO_PRIO_VALUE(IOPRIO_CLASS_BE, 7));
  }
  pthread_setname_np(pthread_self(), "Rvn.Ring.Wrkr");
  struct io_uring *ring = &g_worker.ring;
  struct workitem *work = NULL;
  while (true)
  {
    do
    {
      // wait for any writes on the eventfd 
      // completion on the ring (associated with the eventfd)
      eventfd_t v;
      rc = read(g_worker.eventfd, &v, sizeof(eventfd_t));
    } while (rc < 0 && errno == EINTR);
    
    bool has_work = true;
    while (has_work)
    {
      int must_wait = 0;
      has_work = false;
      if (!work) 
      {
        // we may have _previous_ work to run through
        work = atomic_exchange(&g_worker.head, 0);
      }
      while (work)
      {
        has_work = true;


        struct io_uring_sqe *sqe = io_uring_get_sqe(ring);
        if (sqe == NULL)
        {
          must_wait = 1;
          goto sumbit_and_wait; // will retry
        }
        io_uring_sqe_set_data(sqe, work);
        switch (work->type)
        {
        case workitem_fsync:
          io_uring_prep_fsync(sqe, work->filefd, IORING_FSYNC_DATASYNC);
          break;
        case workitem_write:
          io_uring_prep_writev(sqe, work->filefd, work->op.write.iovecs,
                               work->op.write.iovecs_count, work->offset);
          break;
        default:
          break;
        }
        work = work->next;
      }
    sumbit_and_wait:
      rc = must_wait ? 
        io_uring_submit_and_wait(ring, must_wait) : 
        io_uring_submit(ring);
      struct io_uring_cqe *cqe;
      uint32_t head = 0;
      uint32_t i = 0;


      io_uring_for_each_cqe(ring, head, cqe)
      {
        i++;
        // force another run of the inner loop, 
        // to ensure that we call io_uring_submit again
        has_work = true; 
        struct workitem *cur = io_uring_cqe_get_data(cqe);
        if (!cur)
        {
          // can be null if it is:
          // *  a notification about eventfd write
          continue;
        }
        switch (cur->type)
        {
        case workitem_fsync:
          notify_work_completed(ring, cur);
          break;
        case workitem_write:
          if (/* partial write */)
          {
            // queue again
            continue;
          }
          notify_work_completed(ring, cur);
          break;
        }
      }
      io_uring_cq_advance(ring, i);
    }
  }
  return 0;
}

What does this code do?

We start by checking if we want to use lower-priority I/O, this is because we don’t actually care how long those operations take. The purpose of writing the data to the disk is that it will reach it eventually. RavenDB has two types of writes:

• Journal writes (durable update to the write-ahead log, required to complete a transaction).
• Data flush / Data sync (background updates to the data file, currently buffered in memory, no user is waiting for that)

As such, we are fine with explicitly prioritizing the journal writes (which users are waiting for) in favor of all other operations.

What is this C code? I thought RavenDB was written in C#

RavenDB is written in C#, but for very low-level system details, we found that it is far easier to write a Platform Abstraction Layer to hide system-specific concerns from the rest of the code. That way, we can simply submit the data to write and have the abstraction layer take care of all of that for us. This also ensures that we amortize the cost of PInvoke calls across many operations by submitting a big batch to the C code at once.

After setting the IO priority, we start reading from what is effectively a thread-safe queue. We wait for eventfd() to signal that there is work to do, and then we grab the head of the queue and start running.

The idea is that we fetch items from the queue, then we write those operations to the IO Ring as fast as we can manage. The IO Ring size is limited, however. So we need to handle the case where we have more work for the IO Ring than it can accept. When that happens, we will go to the submit_and_wait label and wait for something to complete.

Note that there is some logic there to handle what is going on when the IO Ring is full. We submit all the work in the ring, wait for an operation to complete, and in the next run, we’ll continue processing from where we left off.

The rest of the code is processing the completed operations and reporting the result back to their origin. This is done using the following function, which I find absolutely hilarious:

int32_t rvn_write_io_ring(
    void *handle,
    struct page_to_write *buffers,
    int32_t count,
    int32_t *detailed_error_code)
{
    int32_t rc = SUCCESS;
    struct handle *handle_ptr = handle;
    if (count == 0)
        return SUCCESS;


    if (pthread_mutex_lock(&handle_ptr->global_state->writes_arena.lock))
    {
        *detailed_error_code = errno;
        return FAIL_MUTEX_LOCK;
    }
    size_t max_req_size = (size_t)count * 
                      (sizeof(struct iovec) + sizeof(struct workitem));
    if (handle_ptr->global_state->writes_arena.arena_size < max_req_size)
    {
        // allocate arena space
    }
    void *buf = handle_ptr->global_state->writes_arena.arena;
    struct workitem *prev = NULL;
    int eventfd = handle_ptr->global_state->writes_arena.eventfd;
    for (int32_t curIdx = 0; curIdx < count; curIdx++)
    {
        int64_t offset = buffers[curIdx].page_num * VORON_PAGE_SIZE;
        int64_t size = (int64_t)buffers[curIdx].count_of_pages *
                       VORON_PAGE_SIZE;
        int64_t after = offset + size;


        struct workitem *work = buf;
        *work = (struct workitem){
            .op.write.iovecs_count = 1,
            .op.write.iovecs = buf + sizeof(struct workitem),
            .completed = 0,
            .type = workitem_write,
            .filefd = handle_ptr->file_fd,
            .offset = offset,
            .errored = false,
            .result = 0,
            .prev = prev,
            .notifyfd = eventfd,
        };
        prev = work;
        work->op.write.iovecs[0] = (struct iovec){
            .iov_len = size, 
            .iov_base = buffers[curIdx].ptr
        };
        buf += sizeof(struct workitem) + sizeof(struct iovec);


        for (size_t nextIndex = curIdx + 1; 
            nextIndex < count && work->op.write.iovecs_count < IOV_MAX; 
            nextIndex++)
        {
            int64_t dest = buffers[nextIndex].page_num * VORON_PAGE_SIZE;
            if (after != dest)
                break;


            size = (int64_t)buffers[nextIndex].count_of_pages *
                              VORON_PAGE_SIZE;
            after = dest + size;
            work->op.write.iovecs[work->op.write.iovecs_count++] = 
                (struct iovec){
                .iov_base = buffers[nextIndex].ptr,
                .iov_len = size,
            };
            curIdx++;
            buf += sizeof(struct iovec);
        }
        queue_work(work);
    }
    rc = wait_for_work_completion(handle_ptr, prev, eventfd, 
detailed_error_code);
    pthread_mutex_unlock(&handle_ptr->global_state->writes_arena.lock)
    return rc;
}

Remember that when we submit writes to the data file, we must wait until they are all done. The async nature of IO Ring is meant to help us push the writes to the OS as soon as possible, as well as push writes to multiple separate files at once. For that reason, we use anothereventfd() to wait (as the submitter) for the IO Ring to complete the operation. I love the code above because it is actually using the IO Ring itself to do the work we need to do here, saving us an actual system call in most cases.

Here is how we submit the work to the worker thread:

void queue_work(struct workitem *work)
{
    struct workitem *head = atomic_load(&g_worker.head);
    do
    {
        work->next = head;
    } while (!atomic_compare_exchange_weak(&g_worker.head, &head, work));
}

This function handles the submission of a set of pages to write to a file. Note that we protect against concurrent work on the same file. That isn’t actually needed since the caller code already handles that, but an uncontended lock is cheap, and it means that I don’t need to think about concurrency or worry about changes in the caller code in the future.

We ensure that we have sufficient buffer space, and then we create a work item. A work item is a single write to the file at a given location. However, we are using vectored writes, so we’ll merge writes to the consecutive pages into a single write operation. That is the purpose of the huge for loop in the code. The pages arrive already sorted, so we just need to do a single scan & merge for this.

Pay attention to the fact that the struct workitem actually belongs to two different linked lists. We have the next pointer, which is used to send work to the worker thread, and the prev pointer, which is used to iterate over the entire set of operations we submitted on completion (we’ll cover this in a bit).

Queuing work is done using the following method:

int32_t
wait_for_work_completion(struct handle *handle_ptr, 
    struct workitem *prev, 
    int eventfd, 
    int32_t *detailed_error_code)
{
    // wake worker thread
    eventfd_write(g_worker.eventfd, 1);
    
    bool all_done = false;
    while (!all_done)
    {
        all_done = true;
        *detailed_error_code = 0;


        eventfd_t v;
        int rc = read(eventfd, &v, sizeof(eventfd_t));
        struct workitem *work = prev;
        while (work)
        {
            all_done &= atomic_load(&work->completed);
            work = work->prev;
        }
    }
    return SUCCESS;
}

The idea is pretty simple. We first wake the worker thread by writing to its eventfd(), and then we wait on our own eventfd() for the worker to signal us that (at least some) of the work is done.

Note that we handle the submission of multiple work items by iterating over them in reverse order, using the prev pointer. Only when all the work is done can we return to our caller.

The end result of all this behavior is that we have a completely new way to deal with background I/O operations (remember, journal writes are handled differently). We can control both the volume of load we put on the system by adjusting the size of the IO Ring as well as changing its priority.

The fact that we have a single global IO Ring means that we can get much better usage out of the worker thread pool that IO Ring utilizes. We also give the OS a lot more opportunities to optimize RavenDB’s I/O.

The code in this post shows the Linux implementation, but RavenDB also supports IO Ring on Windows if you are running a recent edition.

We aren’t done yet, mind, I still have more exciting things to tell you about how RavenDB 7.1 is optimizing writes and overall performance. In the next post, we’ll discuss what I call the High Occupancy Lane vs. Critical Lane for I/O and its impact on our performance.

When we build a new feature in RavenDB, we either have at least some idea about what we want to build or we are doing something that is pure speculation. In either case, we will usually spend only a short amount of time trying to plan ahead.

A good example of that can be found in my RavenDB 7.1 I/O posts, which cover about 6+ months of work for a major overhaul of the system. That was done mostly as a series of discussions between team members, guidance from the profiler, and our experience, seeing where the path would lead us. In that case, it led us to a five-fold performance improvement (and we’ll do better still by the time we are done there).

That particular set of changes is one of the more complex and hard-to-execute changes we have made in RavenDB over the past 5 years or so. It touched a lot of code, it changed a lot of stuff, and it was done without any real upfront design. There wasn’t much point in designing, we knew what we wanted to do (get things faster), and the way forward was to remove obstacles until we were fast enough or ran out of time.

I re-read the last couple of paragraphs, and it may look like cowboy coding, but that is very much not the case. There is a process there, it is just not something we would find valuable to put down as a formal design document. The key here is that we have both a good understanding of what we are doing and what needs to be done.

RavenDB 4.0 design document

The design document we created for RavenDB 4.0 is probably the most important one in the project’s history. I just went through it again, it is over 20 pages of notes and details that discuss the current state of RavenDB at the time (written in 2015) and ideas about how to move forward.

It is interesting because I remember writing this document. And then we set out to actually make it happen, that wasn’t a minor update. It took close to three years to complete the process, to give you some context about the complexity and scale of the task.

To give some further context, here is an image from that document:

And here is the sharding feature in RavenDB right now:

This feature is called prefixed sharding in our documentation. It is the direct descendant of the image from the original 4.0 design document. We shipped that feature sometime last year. So we are talking about 10 years from “design” to implementation.

I’m using “design” in quotes here because when I go through this v4.0 design document, I can tell you that pretty much nothing that ended up in that document was implemented as envisioned. In fact, most of the things there were abandoned because we found much better ways to do the same thing, or we narrowed the scope so we could actually ship on time.

Comparing the design document to what RavenDB 4.0 ended up being is really interesting, but it is very notable that there isn’t much similarity between the two. And yet that design document was a fundamental part of the process of moving to v4.0.

What Are Design Documents?

A classic design document details the architecture, workflows, and technical approach for a software project before any code is written. It is the roadmap that guides the development process.

For RavenDB, we use them as both a sounding board and a way to lay the foundation for our understanding of the actual task we are trying to accomplish. The idea is not so much to build the design for a particular feature, but to have a good understanding of the problem space and map out various things that could work.

Recent design documents in RavenDB

I’m writing this post because I found myself writing multiple design documents in the past 6 months. More than I have written in years. Now that RavenDB 7.0 is out, most of those are already implemented and available to you. That gives me the chance to compare the design process and the implementation with recent work.

Vector Search & AI Integration for RavenDB

This was written in November 2024. It outlines what we want to achieve at a very high level. Most importantly, it starts by discussing what we won’t be trying to do, rather than what we will. Limiting the scope of the problem can be a huge force multiplier in such cases, especially when dealing with new concepts.

Reading throughout that document, it lays out the external-facing aspect of vector search in RavenDB. You have the vector.search() method in RQL, a discussion on how it works in other systems, and some ideas about vector generation and usage.

It doesn’t cover implementation details or how it will look from the perspective of RavenDB. This is at the level of the API consumer, what we want to achieve, not how we’ll achieve it.

AI Integration with RavenDB

Given that we have vector search, the next step is how to actually get and use it. This design document was a collaborative process, mostly written during and shortly after a big design discussion we had (which lasted for hours).

The idea there was to iron out the overall understanding of everyone about what we want to achieve. We considered things like caching and how it plays into the overall system, there are notes there at the level of what should be the field names.

That work has already been implemented. You can access it through the new AI button in the Studio. Check out this icon on the sidebar:

That was a much smaller task in scope, but you can see how even something that seemed pretty clear changed as we sat down and actually built it. Concepts we didn’t even think to consider were raised, handled, and implemented (without needing another design).

Voron HSNW Design Notes

This design document details our initial approach to building the HSNW implementation inside Voron, the basis for RavenDB’s new vector search capabilities.

That one is really interesting because it is a pure algorithmic implementation, completely internal to our usage (so no external API is needed), and I wrote it after extensive research.

The end result is similar to what I planned, but there are still significant changes.  In fact, pretty much all the actual implementation details are different from the design document. That is both expected and a good thing because it means that once we dove in, we were able to do things in a better way.

Interestingly, this is often the result of other constraints forcing you to do things differently. And then everything rolls down from there.

“If you have a problem, you have a problem. If you have two problems, you have a path for a solution.”

In the case of HSNW, a really complex part of the algorithm is handling deletions. In our implementation, there is a vector, and it has an associated posting list attached to it with all the index entries. That means we can implement deletion simply by emptying the associated posting list. An entire section in the design document (and hours spent pondering) is gone, just like that.

If the design document doesn’t reflect the end result of the system, are they useful?

I would unequivocally state that they are tremendously useful. In fact, they are crucial for us to be able to tackle complex problems. The most important aspect of design documents is that they capture our view of what the problem space is.

Beyond their role in planning, design documents serve another critical purpose: they act as a historical record. They capture the team’s thought process, documenting why certain decisions were made and how challenges were addressed. This is especially valuable for a long-lived project like RavenDB, where future developers may need context to understand the system’s evolution.

Imagine a design document that explores a feature in detail—outlining options, discussing trade-offs, and addressing edge cases like caching or system integrations. The end result may be different, but the design document, the feature documentation (both public and internal), and the issue & commit logs serve to capture the entire process very well.

Sometimes, looking at the road not taken can give you a lot more information than looking at what you did.

I consider design documents to be a very important part of the way we design our software. At the same time, I don’t find them binding, we’ll write the software and see where it leads us in the end.

What are your expectations and experience with writing design documents? I would love to hear additional feedback.

RavenDB is meant to be a self-managing database, one that is able to take care of itself without constant hand-holding from the database administrator. That has been one of our core tenets from the get-go. Today I checked the current state of the codebase and we have roughly 500 configuration options that are available to control various aspects of RavenDB’s behavior.

These two statements are seemingly contradictory, because if we have so many configuration options, how can we even try to be self-managing? And how can a database administrator expect to juggle all of those options?

Database configuration is a really finicky topic. For example, RocksDB’s authors flat-out admit that out loud:

Even we as RocksDB developers don't fully understand the effect of each configuration change. If you want to fully optimize RocksDB for your workload, we recommend experiments and benchmarking.

And indeed, efforts were made to tune RocksDB using deep-learning models because it is that complex.

RavenDB doesn’t take that approach, tuning is something that should work out of the box, managed directly by RavenDB itself. Much of that is achieved by not doing things and carefully arranging that the environment will balance itself out in an optimal fashion. But I’ll talk about the Zen of RavenDB another time.

Today, I want to talk about why we have so many configuration options, the vast majority of which you, as a user, should neither use, care about, nor even know of.

The idea is very simple, deploying a database engine is a Big Deal, and as such, something that users are quite reluctant to do. When we hit a problem and a support call is raised, we need to provide some mechanism for the user to fix things until we can ensure that this behavior is accounted for in the default manner of RavenDB.

I treat the configuration options more as escape hatches that allow me to muddle through stuff than explicit options that an administrator is expected to monitor and manage. Some of those configuration options control whether RavenDB will utilize vectored instructions or the compression algorithm to use over the wire. If you need to touch them, it is amazing that they exist. If you have to deal with them on a regular basis, we need to go back to the drawing board.

For a new feature in RavenDB, I needed to associate each transaction with a source ID. The underlying idea is that I can aggregate transactions from multiple sources in a single location, but I need to be able to distinguish between transactions from A and B.

Luckily, I had the foresight to reserve space in the Transaction Header, I had a whole 16 bytes available for me. Separately, each Voron database (the underlying storage engine that we use) has a unique Guid identifier. And a Guid is 16 bytes… so everything is pretty awesome.

There was just one issue. I needed to be able to read transactions as part of the recovery of the database, but we stored the database ID inside the database itself. I figured out that I could also put a copy of the database ID in the global file header and was able to move forward.

This is part of a much larger change, so I was going full steam ahead when I realized something pretty awful. That database Guid that I was relying on was already being used as the physical identifier of the storage as part of the way RavenDB distributes data. The reason it matters is that under certain circumstances, we may need to change that.

If we change the database ID, we lose the association with the transactions for that database, leading to a whole big mess. I started sketching out a design for figuring out that the database ID has changed, re-writing all the transactions in storage, and… a colleague said: why don’t we use another ID?

It hit me like a ton of bricks. I was using the existing database Guid because it was already there, so it seemed natural to want to reuse it. But there was no benefit in doing that. Instead, it added a lot more complexity because I was adding (many) additional responsibilities to the value that it didn’t have before.

Creating a Guid is pretty easy, after all, and I was able to dedicate one I called Journal ID to this purpose. The existing Database ID is still there, and it is completely unrelated to it. Changing the Database ID has no impact on the Journal ID, so the problem space is radically simplified.

I had to throw away heaps of complexity because of a single comment. I used the Database ID because it was there, never considering having a dedicated value for it. That single suggestion led to a better, simpler design and faster delivery.

It is funny how you can sometimes be so focused on the problem at hand, when a step back will give you a much wider view and a better path to the solution.

I write a transactional database for a living, and the best example of why we want transactions is transferring money between accounts. It is ironic, therefore, that there is no such thing as transactions for money transfers in the real world.

If you care to know why, go back 200 years and consider how a bank would operate in an environment without instant communication. I would actually recommend doing that, it is a great case study in distributed system design. For example, did you know that the Templars used cryptography to send money almost a thousand years ago?

Recently I was reviewing my bank transactions and I found the following surprise. This screenshot is from yesterday (Dec 18), and it looks like a payment that I made is still “stuck in the tubes” two and a half weeks later.

I got in touch with the supplier in question to apologize for the delay. They didn’t understand what I was talking about. Here is what they see when they go to their bank, they got the money.

For fun, look at the number of different dates that you can see in their details.

Also, as of right now, my bank account still shows the money as pending approval (to be sent out from my bank).

I might want to recommend that they use a different database. Or maybe I should just convince the bank to approve the payment by the time of the next invoice and see if I can get a bit of that infinite money glitch.