The final aspect of RavenDB’s x7 jump in indexing performance is the fact that we made it freakishly smart.

During standard operation, most indexes only update when new information comes in, we are usually talking about a small number of documents for every indexing run. The problem is what happens when you have a sudden outpour of documents into RavenDB? For example, during nightly ETL batch, or just if you suddenly have a flood of users doing write operations.

The problem here is that we actually have to balance a lot of variable at the same time:

• The number of documents that we have to index*.
• The current memory utilization**.
• How any cores I have available to do the index work with?
• How much time do I have to do this?

Basically, the idea goes like this, if I have a small batch size, I am able to index more quickly, ensuring that we have fresher results. If I have big batch size, I am able to index more documents, and my overall indexing times goes down.

There is a non trivial cost associated with every indexing run, so reducing the number of indexing run is good, but the more documents I shove into a single run, the more memory will I use, and the more time it will take before the results are visible to the users.

* It is non trivial because there is no easy way for us to even know how many documents we have left to index (to find out is costly).

** Memory utilization is hard to figure out in a managed world. I don’t actually have a way to know how much memory I am using for indexing and how much for other stuff, and there is no real way to say “free the memory from the last indexing run”, or even estimate how much memory that took.

What we have decided on doing is to start from a very small (low hundreds) indexing batch size, and see what is actually going on live. If we see that we have more documents to index than the current batch size, we will slowly double the size of the batch. Slowly, because bigger batches requires more memory, and we also have to take into account current utilization, memory usage, and a bunch of other factors as well. We also go the other way around, able to reduce the indexing batch size on demand based on how much work we have to do right now.

We also provide an upper limit, because at some point it make sense to just do a big batch and make the indexing results visible than to try to do everything all at once.

The fun part in all of that is that once we have found the appropriate algorithm for this, it means that RavenDB will automatically adjust itself based on real production load. If you have an low update rate, it will favor small indexing batches and immediately execute indexing on the new documents. However, if you suddenly have a spike in traffic and the update rate goes up, RavenDB will adjust the indexing batch size so it will be able to keep up with your rate.

We have done some (read, a huge amount) testing with regards to this new optimization, and it turns out that under slow update frequency, we are seeing an average of 15 – 25 ms between a document update and it showing up in the indexes. That is pretty good, but what is going on when we have data just pouring in?

We tested this with a 3 million documents and 3 indexes. And it turn out that under this scenario, where we are trying to shove data into RavenDB as fast as it can accept it, we do see an increase in index latency. Under those condition, latency rose all the way to 1.5 seconds.

This is actually something that I am very happy about, because we were able to automatically adjust to the changing conditions, and were still able to index things at a reasonable rate (note that under this scenario, the batch size was usually 8 – 16 thousands documents, vs. the 128 – 256 that it is normally).

Because we were able to adjust the batch size on the fly, we could handle sustained writes at this rate with no interruption in service and no real need to think about this from the users perspective.. Exactly what the RavenDB philosophy calls for.

As I noted in my previous post, we have done major optimizations for RavenDB. One of the areas where we improved the performance was reading the documents from the disk for indexing.

In Pseudo Code, it looks like this:

while database_is_running:
  stale = find_stale_indexes()
  lastIndexedEtag = find_last_indexed_etag(stale)
  docs_to_index = get_documents_since(lastIndexedEtag, batch_size)
  

As it turned out, we had a major optimization option here, because of the way the data is actually structured on disk. In simple terms, we have an on disk index that lists the documents in the order in which they were updated, and then we have the actual documents themselves, which may be anywhere on the disk.

Instead of loading the documents in the orders in which they were modified, we decided to try something different. We first query the information we need to find the document on disk from the index, then we sort them based on the optimal access pattern, to reduce disk movement and ensure that we have as sequential reads as possible. Then we take those results in memory and sort them based on their last update time again.

This seems to be a perfectly obvious thing to do, assuming that you are aware of such things, but it is actually something that is very easy not to notice. The end result is quite promising, and it contributed to the 7+ times improvements in perf that we had for indexing costs.

But surprisingly, it wasn’t the major factor, I’ll discuss a huge perf boost in this area tomorrow.

One of the major dangers in doing perf work is that you have a scenario, and you optimize the hell out of that scenario. It is actually pretty easy to do without even noticing it. The problem is that when you do things like that, you are likely to be optimizing a single scenario to perform really well, but you are hurting the overall system performance.

In this example, we have moved heaven and earth to make sure that we are indexing things as fast as possible, and we tested with 3 indexes, on an 4 cores machine. As it turned out, we actually had improved things, for that particular scenario.

Using the same test case on a single core machine was suddenly far more heavy weight, because we were pushing a lot of work at the same time. More than the machine could process. The end result was that it actually got there, but much more slowly than if we would have run things sequentially.

Of course, I give you the outliers, but those are good indicators for what we found out. Initially, we thought that we could resolve that by using the TPL’s MaxDegreeOfParallelism, but it turned out to be more complex than that. We have IO bound and we have CPU bound tasks that we need to execute, and trying to execute IO heavy tasks with this would actually cause issues in this scenario.

We had to manually throttle things ourselves, both to ensure limited number of parallel work, and because we have a lot more information about the actual tasks than the TPL have. We can schedule them in a way that is far more efficient because we can tell what is actually going on.

The end result is that we are actually using less parallelism, overall, but in a more efficient manner.

In my next post, I’ll discuss the auto batch tuning support, which allows us to do some really amazing things from the point of view of system performance.

One of the things that we are doing during the index process for RavenDB is applying triggers and deciding what, if and how a document will be indexed. The actual process is a bit more involved, because we have to do additional things (like figure out which indexes have already indexed those particular documents).

At any rate, the interesting thing is that this is a process which is pretty basic:

for doc in docs:
    matchingIndexes = FindIndexesFor(doc)
    if matchingIndexes.Count > 0:
       doc = ExecuteTriggers(doc) 
       if doc != null:
          yield doc

The interesting thing about this is that this is a set of operations that only works on a single document at a time, and the result is the modified documents.

We were able to gain significant perf boost by simply moving to a Parallel.ForEach call.  This seems simple enough, right? Parallelize the work, get better benefits.

Except that there are issues with this as well, which I’ll touch on my next post.

The actual process done by RavenDB to index documents is a fairly complex one. In order to understand what exactly happened, I decided to break it apart to pseudo code.

It looks something like this:

while database_is_running:
  stale = find_stale_indexes()
  lastIndexedEtag = find_last_indexed_etag(stale)
  docs_to_index = get_documents_since(lastIndexedEtag, batch_size)
  
  filtered_docs = execute_read_filters(docs_to_index)
  
  indexing_work = []
  
  for index in stale:
    
    index_docs = select_matching_docs(index, filtered_docs)
    
    if index_docs.empty:
      set_indexed(index, lastIndexedEtag)
    else
      indexing_work.add(index, index_docs)
      
  for work in indexing_work:
  
     work.index(work.index_docs)

And now let me show you the areas in which we did some perf work:

while database_is_running:
  stale = find_stale_indexes()
  lastIndexedEtag = find_last_indexed_etag(stale)
  docs_to_index = get_documents_since(lastIndexedEtag, batch_size)
  
  filtered_docs = execute_read_filters(docs_to_index)
  
  indexing_work = []
  
  for index in stale:
    
    index_docs = select_matching_docs(index, filtered_docs)
    
    if index_docs.empty:
      set_indexed(index, lastIndexedEtag)
    else
      indexing_work.add(index, index_docs)
      
  for work in indexing_work:
  
     work.index(work.index_docs)

All of which gives us a major boost in the system performance. I’ll discuss each part of that work in detail, don’t worry