nimdeque/README.md

# nimdeque

Various deque implementations in pure nim. A deque (short for "double-ended queue") is a data type
that is optimized for access towards its ends. A deque's most interesting feauture is the ~O(1)
time that it takes to pop/append at either ends (as opposed to regular lists where appending at the
beginning is an O(n) operation).


------------------------
## Examples

### LinkedDeque

A `LinkedDeque` is a deque based on a doubly linked list.

```nim
import nimdeque


queue = newLinkedDeque[int]()

# Appends at the end
queue.add(1)
queue.add(2)
queue.add(3)

# Prepends at the beginning
queue.addLeft(0)
queue.addLeft(-1)
queue.addLeft(-2)

# Pops the first element in O(1) time
queue.pop()

# Pops the last element in O(1) time
queue.pop(queue.high())  
# This can also be written as
queue.pop(^1)

# Pops element at position n
queue.pop(n)

# Supports iteration
for i, e in queue:
    echo i, " ", e

# Reversed iteration too!
for e in queue.reversed():
    echo e

echo queue.len()
echo 5 in queue  # false
echo 0 in queue  # true

# Item accessing works just like regular sequence types in Nim.
# Note that the further the item is from either end of the
# queue, the higher the time it takes to retrieve it. For
# fast random access, seqs should be used instead
assert queue[0] == -1
assert queue[^1] == queue[queue.high()]

# It's possible to extend a deque with other deques or with seqs
# of compatible type
var other = newLinkedDeque[int]()
other.add(9)
other.add(10)
queue.extend(@[5, 6, 7, 8])
queue.extend(other)

# Clears the queue in O(1) time
queue.clear()
# Clears the queue in O(n) time
queue.clearPop()
```

---------------------
## Notes

- All queue constructors take an optional `maxSize` argument which limits the size of the queue. The
  default value is 0 (no size limit). When `maxSize > 0`, the queue will discard elements from the head when
  items are added at the end and conversely pop items at the end when one is added at the head. Calling `insert`
  on a full queue will raise an `IndexDefect`
- Two deques compare equal if they have the same elements inside them, in the same order. The value of `maxSize` is
  disregarded in comparisons
- Calls to `extend()` **do not** raise any errors when the queue is full. They're merely an abstraction over a for
  loop calling `self.add()` with every item from the other iterable
- Deques in this module do not support slicing. Use the built-in `seq` type if you need fast random accessing and/or slicing
  capabilities
- The objects in this module are **all** tracked references! (Unlike the `std/deques` module which implements them as value
  types and gives `var` variants of each procedure)

## Disclaimer

This is mostly a toy, there are no performance guarantees nor particular optimizations other than very obvious ones. With
that said, the collections _do_ work and are tested somewhat thoroughly (please report any bugs!). The tests directory contains
some benchmarks as well as the test suite used to validate the behavior of the queues.

## Why? There's std/deques!

1. I was bored during my programming class
2. That only provides a deque based on `seq`s
3. The deque in that module is a value type
4. I was bored during my programming class


## Performance metrics
------------------------------------------

To outline just how much faster a deque is than a regular dynamic array when doing insertion and popping
near the ends, here are some performance metrics I took on my laptop with an 8-core Ryzen 7 4700U.

Both collections have been filled with 500000 elements and the time taken for each operation has been
recorded. The average, minimum and maximum time taken are shown here, as well as the standard deviation
for each micro-benchmark.


## Appending at the end

### LinkedDeque

Time taken: `0.325591755` seconds. Results:

  - min:   `0.0`
  - max:   `0.010444388`
  - avg:   `3.415999800000153e-07`
  - stdev: `1.643969031599315e-05`

### seq

Time taken: `0.283606824` seconds. Results:

  - min:   `2.499999999239222e-07`
  - max:   `0.001853333999999984`
  - avg:   `2.750403260000028e-07`
  - stdev: `3.929418946682074e-06`

## Popping the head

### LinkedDeque

Time taken: `0.008184640000000076` seconds. Results:

  - min:   `2.499999999239222e-07`
  - max:   `1.186499999994428e-05`
  - avg:   `2.64712133333553e-07`
  - stdev: `9.675646231323721e-08`

### seq

Time taken: `2.3356288` seconds. Results:

  - min:   `0.0001166700000001519`
  - max:   `0.002396426000000229`
  - avg:   `0.0001549421693333322`
  - stdev: `8.55174518472766e-05`

## Adding at the left side

### LinkedDeque

Time taken: `0.3028111230000001` seconds. Results:

  - min:   `2.699999996913505e-07`
  - max:   `0.01986077199999992`
  - avg:   `3.19837340000251e-07`
  - stdev: `2.808711098829249e-05`

### seq

Time taken: `23.515489255` seconds. Results:

  - min:   `2.700000001354397e-07`
  - max:   `0.002371210999999818`
  - avg:   `4.657242209800181e-05`
  - stdev: `3.14827288610087e-05`

## Random access (10000 times)

### LinkedDeque

Time taken: `8.208724628999995` seconds. Results:

  - min:   `9.920000003660334e-07`
  - max:   `0.002548661999998814`
  - avg:   `0.0008192961785821358`
  - stdev: `0.0004795569727666707`

### seq

Time taken: `0.00333773400000581` seconds. Results:

  - min:   `1.599999990276046e-07`
  - max:   `9.759999997527302e-06`
  - avg:   `1.632506749111258e-07`
  - stdev: `9.612951140246536e-08`

## Popping the tail

### LinkedDeque

Time taken: `0.004907793000000993` seconds. Results:
  - min:   `1.499999982002009e-07`
  - max:   '8.515999994074264e-06`
  - avg:   `1.609196000041869e-07`
  - stdev: `9.069790474947947e-08`

### seq

Time taken: `0.004914629999994702` seconds. Results:
  - min:   `1.499999982002009e-07`
  - max:   `6.697999999971671e-06`
  - avg:   `1.62365333385613e-07`
  - stdev: `5.410771724203826e-08`


## TODOs

There are many possible implementations for double-ended queues: the current one is based on the usual textbook implementation of
a doubly linked list, but that isn't the best choice for cache locality and has significant memory overhead for each link in
the chain; Other possibilities involve using a list of subarrays to alleviate both of these issues, while some other options 
make use of ring buffers or specialized dynamic arrays growing from the center that can be used to allow even fast random accessing 
and can be made really efficient using lazy evaluation. The goal of this module is to implement most (possibly all) of these approaches, 
because I find them fascinating.