A customer was experiencing large memory spikes in some cases, and we were looking into the allocation patterns of some of the queries that were involved. One of the things that popped up was a query that allocated just under 30GB of managed memory during its processing.

Let me repeat that, because it bears repeating. That query allocated 30(!) GB(!) during its execution. Now, that doesn’t mean that it was consuming 30 GB, it was just the allocations involved. Most of that memory was immediately discarded during the operation. But 30 GB of garbage to cleanup puts a lot of pressure on the system. We took a closer look at the offensive query. It looked something like this:

from index “Notifications/RoutingAndPriority”
where startsWith(Route, $routeKeyPrefix)
order by Priority desc

That does not seem like a query that should be all that expensive. But details matter, so we dove into this. For this particular query, the routes are hierarchical structures that are unique for each message. Something like:

• notifications/traffic/new-york-city/67a81019-941b-4d04-a0db-0559ed45343c
• notifications/emergency/las-vegas/0a8e18fb-563b-4b6a-8e93-e10e08239656

And the queries that were generated were using the city & topic to filter the information that they were interested in.

The customer in question had a lot of notifications going on at all times. And each one of their Routes was unique. Internally, RavenDB uses Lucene (currently ) to handle searches, and Lucene is using an inverse index to execute queries.

The usual way to think about is like this:

We have a list of terms (Brown, Green & Purple) and each of them has a list of the matching documents that contain the particular term.

The process of issuing a prefix query then is easy, scan all entries that match the prefix and return their results. This is indeed what Lucene is doing. However… while it is doing that, it will do something like this:

Pay close attention to what is actually happening here. There are two enumerators that we work with. One for the terms for the field and one for the documents for a specific term.

All of this is perfectly reasonable, but there is an issue. What happens when you have a lot of unique values? Well, then Lucene will have a lot of iterations of the loop. In this case, each term has just a single match, and Lucene is pretty good at optimizing search by specific term.

The actual problem is that Lucene allocates a string instance for each term. If we have 30 million notifications for New York’s traffic, that means that we’ll allocate 30 million strings during the processing of the query. We aren’t retaining these strings, mind. They’ll be cleaned up by the GC quickly enough, but that is an additional cost that we don’t actually want.

Luckily, in this case, there is a much simple solution. Given that the pattern of the route is known, we can skip the unique portion of the route. That means that in our index, we’ll do something similar to:

Route = doc.Route.Substring(0, doc.Route.LastIndexOf('/') + 1)

Once that is done, the number of unique matches there would be negligible. There would be no more allocations galore to observe and overall system performance is much improved.

We looked into whether there is something that we can do with Lucene to avoid this allocations issue, but it is endemic to the way the API works. The longer term plan is to fix that completely, of course. We are making great strides there already .

In short, if you are doing startsWith() queries or similar, pay attention to the number of unique terms that you have to go through. A simple optimization on the index like the one above can bring quite a bit of dividends.

I want to comment on the following tweet:

A rather banal “thought experiment:”

What if the only code we could review was tests and interface definitions?

What would that force us to specify at the interface and behaviour level, rather than just the implementation level?

— Reginald Braithwaite (@raganwald) January 2, 2021

When I read it, I had an immediate and visceral reaction. Because this is one of those things that sound nice, but is actually a horrible dystopian trap. It confused two very important concepts and put them in the wrong order, resulting in utter chaos.

The two ideas are “writing tests” and “producing high quality code”. And they are usually expressed in something like this:

We write tests in order to product high quality code.

Proper tests ensure that you can make forward progress without having regressions. They are a tool you use to ensure a certain level of quality as you move forward. If you assume that the target is the tests and that you’ll have high quality code because of that, however, you end up in weird places. For example, take a look at the following set of stairs. They aren’t going anywhere, and aside from being decorative, serves no purpose.

When you start considering tests themselves to be the goal, instead of a means to achieve it, you end up with decorative tests. They add to your budget and make it harder to change things, but don’t benefit the project.

There are a lot of things that you are looking for in code review that shouldn’t be in the tests. For example, consider the error handling strategy for the code. We have an invariant that says that exceptions may no escape our code when running in a background thread. That is because this will kill the process. How would you express something like that in a test? Because you end up with an error raised from a write to a file that happens when the disk is full that kills the server.

Another example is a critical piece of code that needs to be safely handle out of memory exceptions. You can test for that, sure, but it is hard and expensive. It also tend to freeze your design and implementation, because you are now testing implementation concerns and that make it very hard to change your code.

Reviewing code for performance pitfalls is also another major consideration. How do you police allocations using a test? And do that without killing your productivity? For fun, the following code allocates:

Console.WriteLine(1_000_000);

There are ways to monitor and track these kind of things, for sure, but they are very badly suited for repeatable tests.

Then there are other things that you’ll cover in the review, more tangible items. For example, the quality of error messages you raise, or the logging output.

I’m not saying that you can’t write tests for those. I’m saying that you shouldn’t. That is something that you want to be able to change and modify quickly, because you may realize that you want to add more information in a certain scenario. Freezing the behavior using tests just means that you have more work to do when you need to make the changes. And reviewing just test code is an even bigger problem when you consider that you need to consider interactions between features and their impact on one another. Not in terms of correctness, you absolutely need to test that, but in terms of behavior.

The interleaving of internal tasks inside of RavenDB was careful designed to ensure that we’ll be biased in favor of user facing operations, starving background operations if needed. At the same time, it means that we need to ensure that we’ll give the background tasks time to run. I can’t really think about how you would write a test for something like that without basically setting in stone the manner in which we make that determination. That is something that I explicitly don’t want to do, it will make changing how and what we do harder. But that is something that I absolutely consider in code reviews.

In my last post, I looked into the infrastructure of mimalloc, to be frank. A few details on how it is actually hooking the system allocator when used, mostly. I’m still using what is effectively a random scatter/gather approach to the codebase. Mostly because it is small. I’m trying to… grok it would be the best term, I guess. I’m going over the code because it also give me a good idea about practices in what seems to be a damn good C codebase.

I should mention that I drove right into the code, there is also the tech report, which I intend to read, but only after I got through enough of the code to get a good feeling for it.

I run into the code in the options.c file, for instance:

This is a really nice way to get configuration values from the user. What I find really interesting, to be frank, is not the actual options, which would be interesting later on, but the fact that this is such a nice way to represent things in a readable manner.

I’m doing similar things in C# (a much higher level language) to create readable options (typically using dictionary & dictionary initializer). I like how the code is able to express things so succinctly in a language with far fewer features.

However, the order of parameters is critical (is should match the mi_option_t enum values), and there is no way to express this in the code.

I also liked this code, which is part of reading configuration values from env variables:

I like that this is using strstr() in reverse in this manner. It is really elegant.

Going over the OS abstraction layer, on the other hand, show some granliness, take a look here:

I actually simplified the code abit, because it also had #if there for BSD, Linux, etc. I find it harder to follow this style, maybe adding indentation would help, but I have had to read this function multiple times, filtering for specific OSes to get it right.

I did find this tidbit, which is interesting:

This is attempting to do allocation with exclusive access. I wonder how this is actually used for. It looks like mimalloc is attempting to allocate in specific addresses, so that should be interesting.

Indeed, in _mi_os_reset() they will explicitly ask the OS to throw the memory away by calling MADV_FREE or MEM_RESET. I find this interesting, because this let the OS know that the memory can be thrown away, but the allocation still persists. I’m currently looking into some fragmentation issues in 32bits, which won’t be helped by this scenario. Then again, I don’t think that mimalloc is primarily intended for 32 bits systems (I can see code handling 32 bits, but I don’t think this is the primary use case or that 32 bits had a lot of attention).

The mi_os_alloc_aligned_ensured() method call is doing some interesting things. If I need a 16MB buffer, but aligned on 1MB boundary, I have no real way to ask this from the OS. So this is implemented directly by over-allocating. To be fair, I can’t think of a good reason why you’ll want to do something like that (you have no guarantees about what the actual physical memory layout would be after all, and that is the only scenario I can think this would be useful. Given that page aligned memory (which is what you get anyway from the OS) is usually sufficient, I wonder what is the use case for that here.

I get why mimalloc have to struggle with this, given that it is limited to just returning a pointer back (malloc interface), and doesn’t have a good way to tell that it played games with the alignment when you pass a value to free(). There seems to be a lot of code around here to deal with memory alignment requirements.  I think that I’ll need to go up the stack to figure out what kind of alignment requirements it has.

That is enough for now, I think. I’m going to get to the core of mimalloc in the next post.

Unusually for me, I had a bit of a pause in reviewing Sled. As a reminder, Sled is an embedded database engine written in Rust. I last stopped looking at the buffer management, but I still don’t really have a good grasp of what is going on.

The next file is the iterator. It looks like it translates between segments and messages in these segments. Here is the struct:

As you can see, the log iterator holds an iterator of segments, and iterating over it looks like it will go over all the messages in the segments in order. Yep, here is the actual work being done:

The next() method is fairly straightforward, I found. But I have to point out this:

First, the will need call is really interesting. Mostly because you have a pretty obvious way to do conditional compiling that doesn’t really sucks. #if is usually much more jarring in the code.

Second, I think that the style of putting really important functions inside an if result in a pretty dense code. Especially since the if is entered only on error. I would have preferred to have it as a stand alone variable, and then check if it failed.

What I don’t understand is the read_segment call. Inside that method, we have:

There are also similar calls on segment trailer. It looks like we have a single file for the data, but stuff that is too large is held externally, in the blob files.

We then get to this guy, which I find really elegant way to handle all the different states.

That is about it for interesting bits in the iterator, the next fun bit is the Log. I do have to admit that I don’t like the term log. It is too easy to get it confused with a debug log. In Voron, I used the term Journal or Write Ahead Journal (OH in the office: “did we waj it yet?”).

The fact that you need to figure out where to get the offset of the data you are about to write is really interesting. This is the method that does the bulk of the work:

Note that it just reserve and complete the operation. This also does not flush the data to disk. That is handled by the flusher or by explicit call. The reserve() method calls to reserve_internal() and there we find this gem:

I know what it does (conditional compilation), but I find it really hard to follow. Especially because it looks like a mistake, with buf being defined twice. This is actually a case where an #if statement would be better, in my eyes.

Most of the code in there is to manage calls to the iobuf, which I already reviewed. So I’m going to skip ahead and look at something that is going to be more interesting, the page cache. Sled has an interesting behavior, in that it can shred a page into multiple location, requiring some logic to bring it all back together. That is going to be really interesting to look at, I hope.

The file stats with this:

And this… takes a while to unpack.  Remember that epoch is manual GC pattern for concurrent data structure without GC.

The cached_ptr value is a shared pointer to a Node (inside a lock free stack) that holds a CacheEntry with static lifetime and thread safe to a generic argument that must have static lifetime and be thread safe. And there is a unsigned long there as well.

No idea yet what is going on. But here is the first method on this struct:

That is… a lot. The cache entry is a discriminated union with the following options:

There are some awesome documentation comments here, including full blown sample code that really help understand what is going on in the code.

There seems to be a few key methods that are involved here:

• allocate(val) – create a new page and write an initial value, gets back a page id.
• link(id, val) – make a write to a page id. Which simply write a value out.
• get(id) – read all the values for a page id, and uses a materializer to merge them all to a single value.
• replace(id, val) – set the page id to the new value, removing all the other values it had.

The idea here, as I gather. Is to allow sequential writes to the data, but allow fast reads, mostly by utilizing SSD’s random read feature.

I’m trying to follow the code, but it is a bit complicated. In particular, we have:

This try to allocate either a free page or allocate a new one. One of the things that really mess with me here is that the use of the term Page. I’m using to B+Tree, where a page is literally some section of memory. Here it refers to something more nebulous. Key point here, I don’t see where the size is specified. But given that you can link items to page, that sort of make sense. I just need to get used to Pages != storage.

The fact that all of this is generic also make it hard to follow what is actually going on.  I’m getting lost in here, so I think that I’ll stop for now.

FASTER is an interesting project, with some unique approaches to solving their tasks that I haven’t encountered before. When I initially read the paper about a year or so ago, I was impressed with what they were doing, even I didn’t quite grasp exactly what was going on. After reading the code, this is now much clearer. I don’t remember where I read it, but I remember reading a Googler talking about the difference between Microsoft and Google with regards to publishing technical papers. Google would only publish something after it has been in production for a while (and probably ready to sunset ) while Microsoft would publish papers about software that hasn’t been deployed yet.

The reason I mention this is that FASTER isn’t suitable for production. Not by a long shot. I’m talking about issues such as swallowing errors, writing to the console as an error handling approach, calling sleep(), lack of logging / tracing / visibility into what is going on in the system. In short, FASTER looks like it was produced to support the paper. It is proof of concept / research code, not something that can take and use.

You can see it clearly in the way that the system is designed to be operated. You have dedicated threads that process requests as fast as they possibly can. However, there is no concept of working in any kind of operational environment, you can’t start using FASTER from an ASP.Net MVC app, for example. They models are just too different. I can think of a few ways to build a server using the FASTER model, but they are all pretty awkward and very specialized. This lead to the next issue with the project, it is highly specialized solution.

It isn’t meant for general consumption. In fact, as I can figure out, this is perfect if you have a relatively small working set that you do a lot of operations on. The examples I have seen given are related to tracking ads, which is a great example. If you want to store impressions on an ad, the active ads are going to pretty small, but you are going to have a lot of impressions on them. For other stuff, I don’t see a lot of usage scenarios.

FASTER is limited in the following ways:

• Get / Set / Update only – no way to query
• No support for atomic operations on more than a single key
• Can support fixed length values only
• Crash means data loss (of the most recent 14.4 GB, usually)
• The API is awkward to use, and you need to write a bit of (non trivial) code for each key/val you want to store.
• No support for compaction of data beyond dropping the oldest entries

Some of these issues can be mitigated. For example, compaction can be implemented, and you can force data to be written to disk faster if you want to, but those aren’t in the box and require careful handling.

Now that I have gone over the code, I’m not quite sure what was the point there, to be honest. In terms of performance, you can get about 25% of the achieved performance by just using ConcurrentDictionary in .NET, I’m pretty sure that you can do better by literally just using a concurrent hash map in C++. This isn’t something that you can use as a primary data store, after all, so I wonder why not just keep all the data in memory and be done with it.

I liked the mutable / read only portions of the log, that is certainly a really nice way to do it, and I’m sure that the epoch idea simplified things during the implementation with the ability to not worry about concurrent accesses. However, the code is complex and I’m pretty sure that it is not going to be fun to debug / work with in real world scenarios.

To sum it up, interesting codebase and approaches, but I would caution from using it for real. The perf numbers are to salivate over,  but the manner in which the benchmark was written means that it is not really applicable for any real world scenario.

So far, I ignored anything in FASTER about how the data actually hits the disk. Based on my reading of the code and the paper, here is what I think that is going on. FASTER works in segments and in conjunction with its allocator. When you create a new instance, you have to define what would be the log size. From looking at the code, they seem to be favoring 16GB as the default size of the log. This is passed to PersistentMemoryMalloc, which uses pages of 32MB each to manage the memory. Out of these 16GB, 14.4GB are considered to be mutable and 1.6 GB is considered to be read only.

On startup, this class allocates two pages (64MB). Whenever FASTER needs more memory, such as to store a new record, it will eventually call to this method:

Again we have num_slots that actually means size in bytes, but I’ll leave my pet peeves aside.

You can see that this allocates from tail of the page use Reserve, which does an atomic operation. If we run out of space in the 32MB page, the caller need to call NewPage() to handle the new allocation. This plays together with the buffer management and the epoch. In particular, here is how a new page is allocated in the normal case. Assuming we just started and we consumed 64MB of memory, the new entry will allocate the 3rd page. This will also move the section of read only memory when there is a new page allocated and the number of active pages is beyond 14.4 GB.

What this means, in practice, is that FASTER typically has a 14.4GB range in which all operations are working on purely mutable memory. That means that two impressions on the same ad will end up being very close to simply Interlocked.Increment on the relevant value. This is the key for the performance that FASTER exhibits.

What happens once we start going beyond the 14.4 GB? FASTER will begin to move data to the read only section. In this case, it means that the any new modifications to the data will create a new copy of it in the mutable section.

At the same time, PageAlignedShiftReadOnlyAddress() will also starts an async process of writing these readonly pages to disk. Here is how it works:

If the read_only_address was updated, FASTER calls to BumpCurrentEpoch() and will execute OnPagesMarkedReadOnly() when the Epoch moves beyond this range (this works because then it is guaranteed that no one may mutate this memory).  That method looks like this:

The notion of read_only_address and safe_readonly_only_address is discussed in the paper quite nicely, by the way. AsyncFlushPages() writes the data to disk, as you would expect and updates various in memory structures.

Note that just having the data written to disk doesn’t remove the in memory copy. It is still available in memory until it is pushed back from the log by new data. Now that we understand how the data goes to the log and then to the disk, I want to figure out how the data is written in to the disk itself. Actual writes are handled here:

What you can see is that the destination offset is used to divide the data on disk to sections. Each section is 1GB in size. In other words, the way FASTER works is to write the data in 1 GB segments that are sequential over time.

This also plays into the expiration policy that FASTER employs. Because it uses a logs based system, it accumulate data over time and will need to handle that. The current way it deals with the problem is to just delete old files, this gets rid of the data in roughly time based order, which is suitable for the use case that the paper talks about. Another alternative is to read the old files, and move the still valid entries to the start. That doesn’t seem to be implemented and I think it will be pretty hard to do and likely consume a lot of resources.

I’ll probably have another summary post about FASTER, but that is pretty much it. I’ve ignored other parts (recovery, several state machines used to handle internal state, etc), but they aren’t important to grokking what it is actually doing. It is an interesting codebase, but it feels… incomplete. But I’ll reserve such thoughts to the summary post.

One of the things that FASTER claims is that it is suitable for larger than memory datasets with its hybrid log approach. I briefly went over that code during my review, but I want to dedicate this post to discussing how FASTER work with the disk.

FASTER is using async I/O on Linux & Windows to do its I/O, which present its own challenges, in particular, how do you handle an operation that is going to need to hit the disk ( to read old data). Another thing that I would like to discover is how does it save the data to disk. We’ll start from the reading data part.

Reading in FASTER looks like this:

You pass a context, a callback and serial number. I’m not sure what the serial is about, I think it is about checkpoints, but not sure. You’ll be called with the results once the operation executed.

Here is the core of the Read method:

FASTER first checks the in memory details, and if it isn’t there it is either not found or on disk. This is actually really interesting, because it implies that FASTER keep track of all of the data items purely in memory. Let’s go to InternalRead and figure out how that works. We already looked into most of that in the previous post FindEntry is called to find the entry by it’s hash. FASTER keep all the hashes in memory, even while it is writing entries to disk. This way, it can easily answer if an entry exists or not. Note that as I read things, if FASTER has more than a certain number of values, hash collision rate is going to reach high percentage, requiring often expensive I/O to figure out whatever the value exists.

If the value is in memory, your ReadContext.GetAtomic() method will be called. Here is one such implementation:

Where the value was defined as:

If the value has already been moved to the read only section, it will use the Get() method, instead, as an optimization. If the value is not on the mutable section or the read only section, it is on the disk, in which case we have this code:

The go_async() method just records the status of the operation when we started the async process, it doesn’t actually invoke anything async. That is done in the caller code, Read().

This, in turn, gets us to this piece of code:

Note that the code is full of handling of the current phase of the thread. I’m ignoring all of these for now, they aren’t important.

The next thing to look at is the async I/O request, which gets us to:

We first register the pending I/O, then actually starts to process the async call. Except that not really. AsyncGetFromDisk() isn’t actually doing I/O.  Instead, it seems to be focused on limiting the number of concurrent I/O operations that are going on.

In this case, if there are more than 120 pending I/Os, it will call io_getevents() in Linux and GetQueuedCompletionStatus() and try to process any completed I/O immediately.

ProtectAndDrain is more interesting. It asks the epoch to complete any pending operations. Such actions are the creation of a new segment file or the moving of a segment from memory to disk.

I find it interesting that FASTER chose to stall the thread until all pending I/Os are completed. Given their model, it would make more sense to push such operation into the thread ops and resume process other requests. I guess that they expect this to be either a rare case or using this for throttling overall system load. You might also have noticed the num_records arguments, this is used in the hlog method:

That was confusing, I expected this to be the number of records (as in, how many records are read from disk) but this is the number of bytes to read.

The C++ memory model make async code a bit complex. In particular, if you’ll look at the first code snippet in this post, you’ll see that we pass a stack allocated struct to the Read() method. Because this method can complete in an async manner, what FASTER will do is to perform a deep copy of the data. Combine that with lambda’s capturing state, and I’m pretty sure that this code will cause a very subtle memory corruption in rare cases.

What I think will happen is that we capture by ref the stack variable and in 99% of the cases, we’ll run this in sync mode, meaning that there will be no issues. Only if the value needs to be read from disk will you get an issue. Because that function will already return but you still have the callback (and the captured reference now pointing to something completely different) still active. I think that a C++ developer would recognize this, and the fact that C++ require you to be explicit about your captures make this a lot easier to deal with.

It is important to note that there is no good way to actually handle the async nature of certain operations here. Any code that actually handle the operation need to go in the callback.

Another aspect to remember is that just reading from the disk doesn’t mean that you got the right value. You might have gotten a hash collision:

In other words, if you have a way to generate hash collisions, as soon as the value hits the disk, you are going to be facing with making N disk I/O requests to find if you have the value or not.  Denial of service attacks against hash tables are well known and represent a significant risk of to services.

Next on my list if seeing how FASTER is actually writing the data to disk, but that will be in a separate post.

Continuing my stroll through the FASTER codebase, I want to try to get to the core stuff, not just the periphery. This is a bit hard, because there is a lot of code here and it seems to be hard to find where the real action is happening. I decided to find how FASTER is handling Get and Put operations.   There is something called InternalHashTable inside FasterKv, which seems promising,  so I’m going to follow up on that. Interestingly enough, it shows up as:

So there are some funky stuff here to consider too, but we’ll deal with that later, for now, I want to understand what it is doing with the InternalHashTable. Inside that class, there is the notion of bukcets:

And a bucket is:

This starts to get interesting for me, so let’s dig deeper into HashBucketEntry, where they use the same union trick I talked about in my previous posts, this allow to easily define an atomic version of it with very little effort.

There is also the overflow entry definition, which looks like this:

I got to say, I’m really liking this approach for handling data packing as well as multi threading concerns. It also plays very well with the actual cache line architecture of modern CPUs.

There is also the KeyHash struct, whose heart is this:

So far, this lines us very neatly to how the FASTER paper talks about things. Given a KeyHash, here is how you get it’s bucket:

This simply index into the buckets_ array by taking the (size_ –1) bits from the hash itself. That tells me that the FASTER structure is very sensitive to the choice of hash function that will be used. This SO answer has some amazing detail on analysis of the output of hash functions with different outputs, which can give you some idea about why this matters so much. This post is an in depth discussion of this, as well of why the choice of hash function matters. This is enough for me to stop everything and try to find what kind of hash function is actually being used here.

The user get to define their own hash function, and if they do so badly (for example, use the identity function since they have an 8 bytes value and they need an 8 bytes hash) that is going to be fun. The function that they seem to be using in their examples is this one:

A bit of searching on the interweb didn’t narrow it down to something specific, it may be something that they wrote specifically for that. Given that the paper doesn’t mention this, it doesn’t seem to be something special.

Given that 40343 is a prime, it seems like a pretty common pattern of multiple by a prime with each 16 bits portion of the key. The idea is that the multiplication by prime will spread the bits around. No idea how high the quality of this hash function is, since actual analysis would take at least a few hours. At least at a glance,  it doesn’t seem to be awful, but I do wonder whatever something like FNV-1. In fact, this is very similar, but with different prime and offset and with addition instead of XOR.

The actual InternalHashTable class doesn’t actually do something, there are a few functions around checkpointing and recovery, but they aren’t very interesting at this point. I’m going back to actually working with the has table and looked at how reads work. They end up in the InternalRead method which does the majority of the work inside the FindEntry function. The first thing that happens there is:

The first line is there to handle growing the hash table on the fly and will be ignored for now. As you can see, this basically just gives me the relevant bucket. Here is the next stage, once we have a bucket, we need to find the relevant entry that matches what we are looking for. Here is the code:

You can see that there are a bunch of stuff going on here. First, we run over the known entries in the bucket, trying to find an entry with the same tag. You can see the tentative usage, which is used to sync up between concurrent readers and writers. There may be more items in the bucket than we have space for, so there is also the concept of overflow. This is basically a linked list of 7 items at a time and likely to be pretty fast (frequently already in the higher tiers of the cache for commonly accessed data).

Now that we have the entry, let’s see what is done with it. I’m currently reading through the InternalRead code. Just finding the matching hash doesn’t really help us, we may have hash collisions, this is handled here:

This is the first time that we actually run into the hybrid log (hlog here). But the basic idea is pretty obvious. Either the key match, or we have a pointer to the previous entry. I’m not seeing any handling of complete mismatch, though. I’m pretty sure that this is a bug (the behavior is different in the C# version).

This is enough for now, from here the InternalRead code is starting to do things with the hlog, state machine, etc. I’m going to handle that in a separate post.

In my last post I dove into the Epoch implementation. The Epoch is explained very nicely in the paper, and the code follows the paper pretty closely. The code make sense, but I still lack the proper… feeling for how it is actually being used. The Epoch allows you to register code that will be executed when the epoch is updated, which is the key to how FASTER is making progress, but while I can see that this is being called from the allocators, I haven’t really grokked it yet. I’m going to back to the faster.h file and see what I can glean from there.

Because of the template utilization, it is kinda hard to figure out what exactly is going on, I’m going to look at some of the examples and try to figure out what it is doing there. Here is one instance of this class:

AdId and NumClicks are just two ways to provide operations on 8 bytes keys and values. I like these examples because they provide good use case to talk about FASTER usage.

This code leads me to the FileSystemDisk, which is defined as:

In the FileSystemFile, we have this code:

This is pretty simple, but I was quite amused by this, because this is C# API sneaking up again in the C++ code. There is also this:

I’m not sure what this bundle is, though. I run into this code in the class:

This is… not nice, in my eyes. Based on this code, whoever allocated this instance also allocated a buffer large enough to include more data there. This is fairly common, since you want to work with such data together, but I find it ugly / risky because it means that there are multiple locations that needs to be aware of it. I would like it better if they just passed the pointer explicitly. That would avoid this snippet:

Which I find pretty annoying to read. What is stranger is that to use this, you have to write (bundle_t has been typedef for the FileSystemSegmentBundle):

I get what is going on here, but I just find it really awkward to handle. There are multiple places where you need knowledge of this allocation pattern and I don’t believe that the benefit of placing all of the data together is that important. For that matter, given the importance of not using new explicitly in modern C++, I’m sure that there are other ways to achieve the same goal that would be more natural.

Going through the code, we now have:

I decided to make this post about the file system usage, because there is a lot of pretty complex code here that I would like to understand. I finally figured out what the S is, this is the segment size:

This means that the earlier definition of FasterKv basically defined Segment Size of 1 GB in size. I’m not sure what these segments are, though. I’m pretty sure that this is how they manage time base expiration, but I’m not certain. Following upward from the creation of a new segment, we have WriteAsync, like so:

You can see that the segment number is basically just the file id, and if the file does not already exists, we call OpenSegment on it. Afterward, we call WriteAsync on that specific file. I’ll look into how that work in a bit, this isn’t that interesting at the moment. Right now I want to dig into OpenSegment. I removed some error handling here, but the gist of it is clear.

The actual code also handles threading and errors, which I omitted. You can see that it creates the new files, copying them from the existing value. Then it creates a context that holds the old files and pass it to BumpCurrentEpoch.

When FASTER is sure that no one else is looking at this epoch, it will call the callback and delete / dispose the old files. This is a nice way to ensure consistency. LMDB does something similar with its transactions’ table. So now we know that whenever we write at a 1GB boundary, FASTER will generate a new epoch.

What about the actual writing? Here is what this looks like (the Linux impl):

On Linux, this ends up being:

This is then checked in TryComplete:

This is called FasterKv.CompletePending(), which seems to be called occasionally by FASTER. On Windows, this is using async I/O and callbacks to handle this.

Okay, this is already long enough, but I got a handle on how FASTER is writing to disk, even though I don’t know yet what it is doing with that. I also saw an actual use of Epoch that made sense (clearing old data once no one is looking at that).

After going over the paper and the managed implementation, I’m ready to start with the C++ implementation. I have higher hopes for this code. As I started browsing the C++ code, it occurred to me that the way the C#’s implementation handles dynamic code generation is pretty much how templates in C++ work. I wonder if this was the trigger for that.

The C++ code reads a lot more naturally to me. There are some nice tricks that are used there that are a joy to read. For example, take a look at Address manipulation:

The colon syntax are a way to express bitfields in C. But the real fun part for me is the idea of control_. What is this for? Well, it runs out that in addition to Address, the also defined AtomicAddress, whose implementation need to provide atomic operation on address. This is implemented in the following manner:

I find this a really elegant way to handle this requirement.

Another amusing observation is that almost all the code in FASTER is in .h files, because of the heavy use of templates. I wonder how that affects compilation speed and how that would play in larger projects.

It is in faster.h that we start to get into the interesting bits. I first run into this guy:

This maps pretty closely to what I have actually seen the C# code does, but in C++ it is a much more natural approach that dynamic compilation on the fly as it did in C#.

Next we have the constructor, which looks like this:

The epoch_ field is auto initialized by the compiler and is not shown here. This indicates that FASTER can handle up to 2.1 billion entries in total, which seems to be a strange limit for a data store that is expected to handle hundreds of thousands of operations per second. I’m going to jump around the codebase a bit, because I want to follow exactly what is going on when initializing this class. The first place to look is the epoch. The idea of epoch is described in the paper, so I’m not going to repeat it. The code defines a struct that is 64 bytes in size (cache line sized, to avoid false sharing), this is used to store a thread specific value and is used to maintain most of the invariants of the epoch.

When switching between epochs, there are actions that needs to be run, here is what this looks like in the code.

I must say, this really mess up with my mind, because we have C#’s naming conventions (TryPop, TryPush) in C++ code. It’s like the code couldn’t decide what shape it wanted to be in either language.

The number of threads that can take part is limited by this value:

Right now, this is set to 96, which means that if you need more threads than that, you’ll get a runtime error. This fits nicely with the model FASTER uses of long running threads, but I don’t see how it can play well with actually accepting commands from network / other location.

As part of it’s constructor, this method is called, which actually does the real work of setting up the epoch.

I’m not really sure at this point why it is allocating two additional entries beyond the specified size.

When a thread start running FATER code, it needs to register itself with the Epoch, this is done in the Protect() call.

Going into the Thread class reveals a simple table of values that are used to give ids to the threads that asked to get an id. This is done in this function:

It took me a couple of times of reading the first two lines to understand what is going on here. This is an awesome way to handle a circular buffer scanning. It is very clear and saves a bunch of code ( at the cost of doing mod operation, which can be usually be masked if the value is known at compile time and is a power of 2, which in this case it is not). I’m probably going to use this the next time I need to implement scanning through a ring buffer.

Then we have computing the earliest safe epoch:

The first of these methods is elegant, it does a simple read from the table, reading potentially stale values. This doesn’t matter, because the worst thing that can happen is that we’ll keep a previous epoch for longer than it is required.

The second one reads wrong to me, but I’ll have to dig deeper into the C++ memory model more deeply for this. The problem is that this seems like it is relying on the CPU to update its state somehow. But I don’t see any instruction that would force it to. I think that the set to safe_to_reclaim_epoch (which is std::atomic<uint64_t>) will use memory_order_seq_cst for the operation, but I’m not sure how that would impact reads from the table_.

Also, I want you to pay attention to the variable names here. Private member fields:

Public member fields:

And then we have SpinWaitForSafeToReclaim that uses:

• safe_to_reclaim_epoch – public member field
• safe_to_reclaim_epoch_ – method argument

I’m not sure if this a common practice in C++, but this is really confusing to me. This is enough for now, I’m going to keep going thought the C++ code in my next post. There hasn’t been anything really interesting so far, just digging into the code and getting a feel as to how it is put together.