# Sequence cache with TTL expiry

I've had some fun writing a sequence based cache that is going to be used to maintain a consistent stream withstanding network re-connections.

The idea is that subscribers of a network based stream will receive a sequence number with every value in the stream, which can be sent when connecting to the stream which will cause the server to re-send values after the given sequence number, from the cache.

The cache is only required to hold values for a given duration, which will provide a predictable 'maximum-loss' of values after a given time; while not dropping values in periods of high throughput, unlike some sort of circular buffer.

In my implementation, I only process and remove expired values when values are put into the cache - avoiding some sort of schedule based 'clean up' which I didn't really know where to start in terms of implementation, but would likely be useful.

I would like a second opinion on how I've attempted to implement this, and some ways this could be improved or how this problem could be re-thought out entirely.

/// <summary>
/// Stream cache that uses sequence numbers
/// </summary>
public class StreamSequenceCache<T> where T: class
{
/// <summary>
/// The time a cached entry should be kept for
/// </summary>
public TimeSpan TTL { get; set; }

/// <summary>
/// Used to inject date times to use when testing - defaults to DateTime.Now|UtcNow
/// </summary>
public IDateTimeAccessor DateTimeAccessor { get; set; } = SharedLibrary.DateTimeAccessor.Default;

/// <summary>
/// Used to make sure sequence numbers only increase, keeping the values array ordered by sequence numbers
/// </summary>
long lastSequenceNumber = long.MinValue;

/// <summary>
/// Values in the cache
/// </summary>
List<CacheEntry> values = new List<CacheEntry>();

/// <summary>
/// Put a value into the cache, for a given sequence number
/// </summary>
/// <param name="sequenceNumber"></param>
/// <param name="value"></param>
public void Put(long sequenceNumber, T value)
{
if (sequenceNumber <= lastSequenceNumber)
throw new InvalidOperationException("Sequence number must only be incremented");

var now = DateTimeAccessor.UtcNow;

values.Add(new CacheEntry(sequenceNumber, now, value));
lastSequenceNumber = sequenceNumber;

// Find the first non-expired index - which will be the count to remove all expired values
// preceding it
// The max is to turn -1 (not found) into 0 (remove none)
// Note that -1 could also mean all entries have expired, but that isn't possible since we just inserted a non-expired entry
var removeCount = Math.Max(values.FindIndex(e => e.Timestamp + TTL > now), 0);
values.RemoveRange(0, removeCount);
}

/// <summary>
/// Get the values after a given sequence number
/// </summary>
/// <param name="sequenceNumber"></param>
/// <returns></returns>
public IEnumerable<T> Get(long afterSequenceNumber)
{
try
{
// Find the sequence number to return values after
int index = values.BinarySearch(new CacheEntry(afterSequenceNumber));

// If the index is less than 0, then it is some complementary something I don't understand... but using the magic
// ~ operator I can get back the index it 'would have' been inserted, which is effectively the count to skip
if (index < 0)
index = ~index;
else
// Increment by 1, since we need to skip the value at the index
index++;

// Return values
return values.Skip(index)
.Select(s => s.Value);
}
catch(InvalidOperationException e)
{
// Sequence number not found. We simply ignore it, since it is probably to old to care.
return new List<T>();
}
}

/// <summary>
/// Container to hold a cached entry
/// </summary>
class CacheEntry : IComparable<CacheEntry>
{
/// <summary>
/// Used when binary searching for a sequence number
/// </summary>
/// <param name="sequenceNumber"></param>
public CacheEntry(long sequenceNumber)
{
SequenceNumber = sequenceNumber;
}

public CacheEntry(long sequenceNumber, DateTime timestamp, T value)
{
SequenceNumber = sequenceNumber;
Timestamp = timestamp;
Value = value;
}

public long SequenceNumber { get; set; }
public DateTime Timestamp { get; set; }
public T Value { get; set; }

/// <summary>
/// Compare only the sequence number. This is used when binary searching to find the entries to respond
/// to a get request for
/// </summary>
/// <param name="other"></param>
/// <returns></returns>
public int CompareTo(CacheEntry other)
{
return SequenceNumber.CompareTo(other.SequenceNumber);
}
}
}


Additionally, I have some unit tests which I used to test this class - which some feedback would be very valuable to someone who doesn't write many unit tests:

using Microsoft.VisualStudio.TestTools.UnitTesting;
using SharedLibrary.Caching;
using System;
using System.Linq;
using TestSharedLibary.Mocks;

namespace TestSharedLibary
{
[TestClass]
public class TestStreamSequenceCache
{
MockDateTimeAccessor dateTimeAccessor = new MockDateTimeAccessor();

[TestMethod, Description("Test putting values")]
public void TestPut()
{
var cache = new StreamSequenceCache<string>();
cache.DateTimeAccessor = dateTimeAccessor;

cache.Put(-1, "item 1");
cache.Put(0, "item 2");
cache.Put(1, "item 3");
cache.Put(2, "item 4");

// Make sure dublicate sequence throws an exception
Assert.ThrowsException<InvalidOperationException>(() =>
{
cache.Put(2, "item 5");
});
}

[TestMethod, Description("Test getting sequences from the cache")]
public void TestGet()
{
var cache = new StreamSequenceCache<string>();
cache.DateTimeAccessor = dateTimeAccessor;

cache.Put(0, "item 1");
cache.Put(1, "item 2");
cache.Put(2, "item 3");
cache.Put(3, "item 4");
cache.Put(4, "item 5");

Assert.AreEqual(0, cache.Get(5).Count());
Assert.AreEqual(5, cache.Get(-10).Count());
Assert.AreEqual(2, cache.Get(2).Count());
Assert.AreEqual(0, cache.Get(4).Count());

cache = new StreamSequenceCache<string>();
cache.DateTimeAccessor = dateTimeAccessor;

cache.Put(-5, "item 1");
cache.Put(-4, "item 2");
cache.Put(3, "item 3");
cache.Put(4, "item 4");
cache.Put(5, "item 5");

Assert.AreEqual(4, cache.Get(-5).Count());
Assert.AreEqual(3, cache.Get(-4).Count());
Assert.AreEqual(1, cache.Get(4).Count());
Assert.AreEqual(0, cache.Get(5).Count());
Assert.AreEqual(0, cache.Get(6).Count());
}

[TestMethod, Description("Test cache expiry")]
public void TestExpiry()
{
var cache = new StreamSequenceCache<string>();
cache.TTL = TimeSpan.FromSeconds(5);
cache.DateTimeAccessor = dateTimeAccessor;

var reference = DateTime.UtcNow;

dateTimeAccessor.UtcNow = reference.AddSeconds(0);
cache.Put(0, "item 1");
dateTimeAccessor.UtcNow = reference.AddSeconds(1);
cache.Put(1, "item 2");
dateTimeAccessor.UtcNow = reference.AddSeconds(2);
cache.Put(2, "item 3");
dateTimeAccessor.UtcNow = reference.AddSeconds(3);
cache.Put(3, "item 4");
dateTimeAccessor.UtcNow = reference.AddSeconds(4);
cache.Put(4, "item 5");

// Check all items were added
dateTimeAccessor.UtcNow = DateTime.UtcNow;
Assert.AreEqual(5, cache.Get(-100).Count());

// Check expiring 1 item
dateTimeAccessor.UtcNow = reference.AddSeconds(5.5);
cache.Put(5, "item 6");
// Should still be 5 items, since the last should have been dropped
Assert.AreEqual(5, cache.Get(-100).Count());

// All items should be dropped
dateTimeAccessor.UtcNow = reference.AddSeconds(20);
cache.Put(6, "item 7");
// Should still be 5 items, since the last should have been dropped
Assert.AreEqual(1, cache.Get(-100).Count());
}
}
}


## 1 Answer

I think StreamSequenceCache<T> does not need to be extended, it may be sealed.

From top to bottom. TTL should be Ttl and you should validate the value in the property setter (is a negative value valid?)

Also setter of DateTimeAccessor should validate its input unless a null value is acceptable.

I'd add the private modifier even when it's implicit, but it's pretty opinionated.

lastSequenceNumber should really be ulong instead of long. It may have sense to keep it signed only if you have to interop with languages without unsigned values support (but in that case you may expose an opaque value instead of a primitive type).

I don't think List<T> is the best data structure for your task. I don't have real world data to analyze for your usage scenario but frequently removing items from the top of the list is an expensive operation. Maybe a LinkedList<T> or a custom data structure based on a queue (or an array with a head and tail indices) may work better. Some testing and profiling is necessary here.

In Get() to return an empty list you do not need to create a new empty List<T> object, you can use Enumerable.Empty<T>(). Note that when searching for an index you do not need that magic, you're playing with the fact that ~(-1) is 0 then you effectively skip nothing...I think an explicit check is more clear:

if (index < 0)
return values.Select(x => x.Value);


Also you may consider to write your own BinarySearch() method to avoid creation of a fake CacheEntry object. If performance matters and the list may become pretty big you should also fine tune the algorithm little bit (to favor the tail of your list).

CacheEntry can be sealed. Also you do not need setters for its properties, you may, for example, change them to:

public ulong SequenceNumber { get; }


You should consider to use struct instead of class. It's not an easy decision and you should absolutely profile, also I think struct size is on the edge in this case.

While to use xmldoc is a very good habit, I think you may avoid it for private/internal elements (especially if you do not use it consistently for everything). When you need a short note for developers you may use a normal comment or a <devdoc>.

This class is obviously not thread-safe, I do not see your usage pattern then I can't comment on this.

• I see your point regarding the ulong, but fail to see how that would simplify the Get method at all. Would you mind explaining? – Warrick Aug 23 '17 at 8:33
• And on a side note, should all classes in general be designed to be thread safe? In my situation, the entire application is single threaded to avoid unnecessary complexity and race conditions but I'm just wondering what the expectation is for classes like this. – Warrick Aug 23 '17 at 8:35
• Thread-safe? Absolutely not, you're right to do not design it to be thread-safe if it's not required or it's something else responsibility. I just meant that I don't know how you will use it (it's a server then multiple concurrent requests are, I guess, expected) then I can't say if you're missing something or not. – Adriano Repetti Aug 23 '17 at 8:38
• About Get() you're right, I just pushed keys with my back, let me fix it. – Adriano Repetti Aug 23 '17 at 8:39