Skip to main content

Sort a bulk of data in disk and parallel (RAM memory free)

Project description

Sort in disk

Sort a bulk of data in disk (RAM memory free) and optionally in parallel (Use available resources).

It is recomendable use sorted_in_disk to sort bulk of data to disk, because sorted save data RAM memory that is pretty limited.

This is a good way to work with big data without RAM memory limits (hard disk have more size and cheaper than RAM memory).

sorted_in_disk algorithm return data as soon as posible and only work the minimum necessary, that mean sorted is in real time while return sorted data to you (in injection time perform a quick and minimum sorted work).

Table of contents

Technologies

Project create with:

  • Python 3

Setup

Install it locally using PyPI (last release in https://pypi.org/project/sorted-in-disk/):

pip install sorted-in-disk

Information about dependencies

Mandatory dependencies (auto-import with PIP):

  • easy_binary_file: to manage binary files
  • quick_queue: to pass quick values between processes

Optional dependencies (only hand-installable):

  • psutil: to check current memory of process and save to disk (if you do not have psutil, then when certain number of items will be reached, then save to disk). Not mandatory because I detected some incompatibilities issues in some systems (psutil need extra permissions). Optionally you can install psutil with:
pip install psutil

Quick use

Import:

from sorted_in_disk import sorted_in_disk

Example of inject a list of string lines with pipe | delimiter where key is first part:

unsorted_data = [
    "key3|value3",
    "key1|value1",
    "key2|value2"
]

sid = sorted_in_disk(unsorted_data,
                     key=lambda line: line.split("|")[0])

for element in sid:
    print(element)

You can read from sorted_in_disk like a dict style with values, keys, items methods. Example:

for key, value in sorted_in_disk(...).items():
    pass

SortInDisk implements __iter__ (with values). Example:

for value in sorted_in_disk(...):
    pass

If you have a big text file where each line has a key to sort, you can read file line to line quickly with read_iter_from_file in sorted_in_disk.utils package to pass an iterable for lines in the file (this way not read full file in one time, read line per line in a generator; only consume one line size in RAM memory). Example (key supposes file have lines similar to "key1|value1"):

from sorted_in_disk.utils import read_iter_from_file

iterable_with_unsorted_data = read_iter_from_file("path/to/your/file/to/read")

sid = sorted_in_disk(iterable_with_unsorted_data,
                     key=lambda line: line.split("|")[0])

for sorted_line in sid:
    print(sorted_line)

In comparison with sorted method

sorted_in_disk is similar to oficial sorted method (https://docs.python.org/3/library/functions.html#sorted), with iterable, key and reverse args.

Example for comparison of common sorted use:

unsorted_data = [
    "key3|value3",
    "key1|value1",
    "key2|value2"
]

for element in sorted(unsorted_data,
                      key=lambda line: line.split("|")[0]):
    print(element)

Example of sorted_in_disk use:

for element in sorted_in_disk(unsorted_data,
                              key=lambda line: line.split("|")[0]):
    print(element)

Algorithm lifecycle

sort_in_disk method creates one instance of SortInDisk object.

The sort in disk method have a lifecycle:

  • Clean before: Check if not exist previous structure in disk and clean if it is necessary
  • Injection: Data unsorted is inject in this structure
  • Reader: Data sorted is read from this structure
  • Clean after: clean if it is necessary

Clean before

One live instance of SortInDisk own a temporal dir path to create temporal files.

No more than one instance of SortInDisk at time may own to same temporal dir path.

For this reason is important check if not exist previous temporal files generated for sort_in_disk in same folder in disk.

One instance of `SortInDisk can:

  • Create a new temporal dir: Not exist other temporal dir with same name. This is a clean state.
  • Append more data to previous temporal dir: If exist a temporal dir with same name and not other instance that pointing to this dir are live, then this new instance can own previous temporal dir to append more data. This is enable if sorted_in_disk have arg append to True.
  • Delete previous temporal dir: delete all, similar to clean state and begin. This is enable if sorted_in_disk have arg append to False.
  • Rename a new temporal dir: in case of exist a previous temporal dir with same name, create a new temporal dir with an increment number and use this new as clean state. This is enable if sorted_in_disk have arg ensure_different_dirs to True.

Injection

Data can be injected in multiprocess or mono-process (in the main thread) way.

Each thread to create a new raw file with all injected data.

Also, each thread create a cache in RAM memory with index of keys to point data and extra information.

By other hand, it is created one file more with general information.

If you have psutil installed, then if this cache has more size than max_write_process_size and count_insert_to_check is reached (to reduce check and increase injection speed), then keys cached are pre-sorted and saved to disk and create a new empty cache.

In other way, if you do not have psutil installed (or if max_write_process_size is None), then if count_insert_to_check is reached, then keys cached are pre-sorted and saved to disk and create a new empty cache.

You can enable mono-process if sorted_in_disk have arg write_processes to 0.

You can enable multiprocess if sorted_in_disk have arg write_processes to None to auto-determinate physical processors in current system or define a number of processors. Depend on your velocity of main-thread reading data you take advantage of multiprocess or not. If multiprocess is enable, then when main-thread end of inject data next code will be not blocked; it is necessary join with join_multiprocess(), some methods call to join_multiprocess() as __iter__, values, keys, items or __len__. Multiprocess have a queue determine with queue_max_size to work with data between processes (it can be regulated, but it is better not to put neither too much to have enough RAM memory nor too little to do not have idle process).

To sum up process control:

  • write_processes = 0 is mono-process injection.
  • write_processes = None is multi-process injection, one process per physical processor detected in system.
  • write_processes = 4 is multi-process injection, 4 process to inject data.
  • write_processes = ["path/tmp_process_1", "path/tmp_process_2", "path/tmp_process_3"] is multi-process injection, 3 process to inject data, one per directory.

To sum up memory control:

  • max_write_process_size = None or not psutil, and count_insert_to_check = 1000000 then save index to disk when 1000000 values binjected, clean and continue.
  • max_write_process_size = 1024*1024*1024 or not psutil, and count_insert_to_check = 1000000 then check if 1 GB are reached each 100000 values inject, if precess reach 1 GB, then index save to disk, clean and continue.
  • queue_max_size only if multi-process injection is enable. If queue_max_size=1000 then main process put in the queue max 1000 values, those values will be taking by consumption processes. If data does not fit in the queue then main process will go to idle until the queue have space.

Reader

Data can read in multiprocess or mono-process (in the main thread) way.

In one thread, the header of data read from all files and sorted these headers. Then it return the first header, choose next header from file and sort other time all headers and repeat.

In this point have data in real time (data streaming). Data is not necessary full sorted to sure a correct sort.

If full data read, then file will be closed (and go to after clean step in life cycle).

You can enable mono-process if sorted_in_disk have arg read_process to False.

You can enable multiprocess if sorted_in_disk have arg read_process to True. In this case, data is sorted and enqueue until these sorted data will read in main-thread. This is useful if we want to take advantage of the data preparation between data consumption. Multiprocess have a queue determine with iter_m_queue_max_size to work with data between processes (it can be regulated, but it is better not to put neither too much to have enough RAM memory nor too little to do not have idle process).

To sum up:

  • read_process = False is mono-process reader.
  • read_process = True is multi-process reader (one process more to prepare a bulk of sorted data).
  • iter_m_queue_max_size only if multi-process reader is enable. If iter_m_queue_max_size=1000 then second process have a queue of 1000 positions to put data, if data does not fit in the queue then process will go to idle until queue main process take data from the queue and leave space.

Clean after

Depend on we want append more data or maintain sorted work in disk, maybe we do not want clean anything.

In the end on one instance of `SortInDisk can:

  • Maintain the temporal dir: This not delete temporal files to use other times in the future. If we want clear data in the future, we will be able to delete the temporal folder from disk or use sorted_in_disk other time and change the only_one_read to True or use clear method by hand. To enable maintain the temporal dir if sorted_in_disk have arg only_one_read to False.
  • Delete the temporal dir: delete all temporal data, clean state. This is enable if sorted_in_disk have arg only_one_read to True.

Performance

Hardware

sorted_in_disk take advantage of you Hardware, then this is more quickly if it is well configured and run in best hardware.

You need to know:

  • RAM memory: it is used to accelerate injection, multiprocess queues and pre-sorted, when condition of variables count_insert_to_check and queue_max_size are met, then will dump to disk to liberate RAM memory. In conclusion, if you have more RAM memory you can configure more count_insert_to_check and queue_max_size to work more in faster RAM memory. Example:
sid = sorted_in_disk(...,
                     count_insert_to_check=1000000,
                     max_write_process_size=1024 * 1024 * 1024)
  • Processes: To inject/write is a good idea divide work in processes one per processor (with write_processes variable) if it is possible to increase the velocity of sorted in disk. But, you need to calibrate the best number of processes in your system if you want to reach the max performance; because, maybe more processes are no more quickly if you have RAM memory or disk limitations. To read, if read_process=True you have a dedicated precess to read while you consume data and have a buffer of prepared sorted data.

  • Disk: It is better work in SSD disk than HHD disk to prevent a bottleneck of processes. To increase the disk performance you can read data from different disk (depend on iterable read data) where the sorted is taking place. Configure main path with tmp_dir (or/and define one list with paths to other disks with write_processes). Example:

sid = sorted_in_disk(...,
                     tmp_dir="/path/to/tmp_dir")

To start I recommend you test first with mono-process (write_processes=0) in injection and read (read_process=False).

sid = sorted_in_disk(...,
                     write_processes=0)

After, you can test with multiprocess (example for 4 inject processes write_processes=4, or create one per physical processor with write_processes=None). If you think velocity of disk is enough, you can enable multiprocess and test with different configurations. If disk velocity is not enough then is quickly and easy use mono-process.

sid = sorted_in_disk(...,
                     write_processes=4)

If you think your bottleneck is disk, and you want to enable multiprocess, then you can try to pass a list with paths to write_processes. Each element of this list create a process than save data in each assigned path (tmp_dir continue saving a small file with state, but not bulk of data). For example, you have tree disks, "disk_HDD1" and "disk_HDD2" are HDD and "disk_SSD" is a SSD, then you can define one process per HDD disk and two process to SSD because it is quickly (in total 4 injection process; similar to write_processes=4 but each process have one path defined to save data):

sid = sorted_in_disk(...,
                     tmp_dir="/path/to/tmp_dir",
                     write_processes = ["disk_HDD1/path/tmp_HDD1", 
                                        "disk_SSD/path/tmp_SSD", 
                                        "disk_SSD/path/tmp_SSD", 
                                        "disk_HDD2/path/tmp_HDD2"])

To read, if you need to perform hard operations with each returned data, it is better to enable multi-process in read.

sid = sorted_in_disk(...,
                     read_process=True)

In other case, it is quick enough mono-process.

sid = sorted_in_disk(...,
                     read_process=False)

More advanced performance settings in sorted_in_disk if you use multi-process, then you can configure the queue to read or write (both queues are independents).

For example to precise fit configuration of write-queue (if write_processes=0 write-queue not work):

sid = sorted_in_disk(...,
                     write_processes=None,  # Write multiprocess enabled if this is different to 0 (None, 1, 2, 3, etc...)
                     queue_max_size=1000,
                     size_bucket_list=None,
                     min_size_bucket_list=10,
                     max_size_bucket_list=None)

Or an example to precise fit configuration of read-queue (if read_process=False read-queue not work):

sid = sorted_in_disk(...,
                     read_process=True,  # Read multiprocess enabled if this is True
                     iter_m_queue_max_size=1000,
                     iter_min_size_bucket_list=10,
                     iter_max_size_bucket_list=None)

Reuse pre-sorted work

You can use many times one sorted work from disk (if only_one_read=False), but this is not a data base. Example:

sid = sorted_in_disk(...,
                     only_one_read=False)

When data is in disk have a minimum sorted work, but it is not finally sort. When you read data perform complete sort in real time (to have sorted data as soon as posible). Due to, if you want use several times sorted work, maybe is good idea save result to file and read from this one (if you want to take advantage of same read iteration, with Python generators in streaming configuration you can save data to disk while you use data at same time).

Performance test

Hardware where the tests have been done:

  • Processor: Intel i5 3.2GHz 4-core
  • Operating System: Windows 10 x64
  • RAM Memory: 8 GB

Use different configurations in python3 tests\complex_example.py

Inject N elements in sorted_in_disk.

6,360,077 elements (1 GB of data) <All data sorted in same HHD disk using 1,16 GB>

Mono-Process: main process (injection in same main process):

sorted_in_disk write and pre-sort: 0:01:41.586510

Example to instance this execution:

sid = sorted_in_disk(iterable_1gb,
                     key=lambda line: line.split("|")[2],
                     write_processes=0)

+1 Process: main process and 1 process for injection (time: Mono-Process = 1 Process x 3.3 faster):

sorted_in_disk write and pre-sort: 0:00:30.138000

Example to instance this execution:

sid = sorted_in_disk(iterable_1gb,
                     key=lambda line: line.split("|")[2],
                     write_processes=1)

+4 Processes: main process and 4 processes for injection (time: Mono-Process = 4 Processes x 3.5 faster):

sorted_in_disk write and pre-sort: 0:00:28.697512
sid = sorted_in_disk(iterable_1gb,
                     key=lambda line: line.split("|")[2],
                     write_processes=None)

314,610,000 elements (50 GB of data) <All data sorted in same HHD disk using 66 GB>

Mono-Process: main process (injection in same main process; max 2 GB RAM memory consumption):

sorted_in_disk write and pre-sort: 2:01:26.222081

Example to instance this execution:

sid = sorted_in_disk(iterable_50gb,
                     key=lambda line: line.split("|")[2],
                     write_processes=0)

+4 Processes: main process and 4 processes for injection (max 5 GB RAM memory consumption) (time: Mono-Process = 4 Processes x 2 faster):

sorted_in_disk write and pre-sort: 1:16:38.016659
sid = sorted_in_disk(iterable_50gb,
                     key=lambda line: line.split("|")[2],
                     write_processes=None)

Readed

Reminder: sorted_in_disk write and do pre-sort operations while consume all data, after read in realtime when data is consumed by third code. In the following individual performance tests it is only indicated "write and pre-sort" heavy work require consume all data and block main process until all data will inject.

In all cases, read sorted data in realtime (reading in same main process from all files):

100,000 elements each 4 seconds approximately

To configure read in same process (with block while process of sort improved):

sid = sorted_in_disk(iterable,
                     ...
                     read_process=False)

To configure read in different process (without block while process of sort improved):

sid = sorted_in_disk(iterable,
                     ...
                     read_process=True)

Documentation

This is a review of class and functions content inside.

You have complete docstring documentation in code and more examples/tests in doctest format.

Function:

  • sorted_in_disk: Main method to create a SortedInDisk object configured

Helpers methods public to take advantage of this package (those are not main package use):

  • create_tmp_folder: Helper to create a temporal folder
  • delete_tmp_folder: Delete temporal files created

sorted_in_disk args

Args to configure common sorted args:

  • iterable: iterable to sort in disk. This iterable should by a generator for big data due to RAM memory limitation.
  • key: key specifies a function of one argument that is used to extract a comparison key from each element in iterable (for example, key=str.lower or key=lambda e: e.split(",")[3]). The default value is None (compare the elements directly).
  • value: value specifies a function of one argument that is used to extract a value from each element in iterable (for example, key=lambda e: e.split(",")[1:]).
  • reverse: reverse is a boolean value.

Args to configure temporal directories: If set to True, then the list elements sorted as if each comparison were reversed.

  • tmp_dir: Path to dir where save temporal files. If None this creates a folder and overwrite if exist previously. By default: create a sortInDiskTmps folder in the current directory.
  • ensure_different_dirs: True to add incremental counter to folder if exists previously. Useful if you use several instances of SortedInDisk at same time. Note: conflict if append is True, because this creates a new name and not delete previously file (example if True: if exist /path/folder/ it create a new /paht/folder(1)/). By default: False
  • append: True to clean folder tmp_dir if existe previously. By default: False
  • only_one_read: True to clean folder tmp_dir when you consume all data. If it is True only works if you read all returned data, if you not read all, then you need to clear instance to auto. By default: True

Args to configure write/injection:

  • count_insert_to_check: counter to check if process have more size in memory than max_write_process_size. By default: 1000000
  • max_write_process_size: max size in bytes to dump cache memory values to disk (only execute when count_insert_to_check is reached). If None, then not import psutil and then only check with count_insert_to_check. By default: 1024*1024*1024 # 1Gib
  • ensure_space: True to ensure disk space but is slowly. If not space then process launch warning message and wait for space. If False, then get and IOException if not enough space. By defatul: False
  • write_processes: number of process to execute. If None then it is number of CPUs. If you pass one list with paths pointing to folders, then each path implements one process (each process save data in its own path; you can use one path to several processes if you define same path several times in the list) and these paths are managed by sorted_in_disk (tmp_dir continues to be used for save general state information). By default: 0
  • queue_max_size: (only if write_processes!=0) max number of elements in queue. If None then is the max by default. By default: 1000
  • size_bucket_list: None to enable sensor size bucket list (require maxsize>0). If a number is defined here then use this number to size_bucket_list and disable sensor. If maxsize<=0 and size_bucket_list==None then size_bucket_list is default to 1000; other wise, if maxsize<=0 and size_bucket_list is defined, then use this number. By default: None
  • min_size_bucket_list: (only if sensor is enabled) min size bucket list. Min == 1 and max == max_size_bucket_list - 1. By default: 10
  • max_size_bucket_list: (only if sensor is enabled) max size bucket list. If None is infinite. By defatult: None Args to configure read:
  • read_process: True to get and prepare data in other process, False to use this one. By default: False
  • iter_m_queue_max_size: (only if enable_multiprocessing is True) max number of elements in queue. If None then is the max by default. By default: 1000
  • iter_min_size_bucket_list: (only if sensor is enabled) min size bucket list. Min == 1 and max == iter_max_size_bucket_list - 1. By default: 10
  • iter_max_size_bucket_list: (only if sensor is enabled) max size bucket list. If None is infinite. By default: None Args to debug:
  • logging_level: Level of log. Only to debug or to remove psutil warning. By default: logging.WARNING

Class:

  • SortedInDisk: Instance an object to work with data in a specific temporal folder.
    • save_and_sort: Choose save_and_sort_multiprocess of save_and_sort_mono depend on write_processes
    • __iter__: Sorted iterable of lines (same as values method).
    • __len__: Get number of elements in this structure.
    • items: Get a sorted iterable from disk to return sorted tuples of key and line, in each petition this get one sorted
    • values: Get a sorted iterable from disk to return sorted lines, in each petition this get one sorted line.
    • keys: Get a sorted iterable from disk to return sorted keys of lines, in each petition this get one sorted key.
    • join_multiprocess: Wait to end of all processes (only it is important if multiprocess injection is enable).
    • clear: Clear file and delete temporal files
    • visor: Visor of information in state file.
    • Other methods invoked in previous methods (public for package extension proposals):
      • delete_tmp: Delete temporal files created (use clear to use instance state)
      • get_dict_saved_info: Get dict with general information. If not exist, create a new empty.
      • set_dict_saved_info: Save in disk a new dict with general information.
      • save_and_sort_multiprocess: Consume an iterable to be sorted in multiprocess way. Take analysis in this iterable and save to disk (in temporal files).
      • save_and_sort_mono: Consume an iterable to be sorted. Take analysis in this iterable and save to disk (in temporal files). Mono-thread, this one execute in the current thread.

Utils functions:

Some tools to make work easier to read a file from disk to use sorted_in_disk and others.

  • write_iter_in_file: Write a iterable as text line in file
  • read_iter_from_file: Read a iterable where each element is a text line in file
  • human_size: Return a human size readable from bytes

How to read a file and sort quickly

You have a file similar to this content, and you want to sort with "keyN":

valA,key3,valB
valC,key1,valD
valE,key2,valF

You can inject use util read_iter_from_file in this way:

from sorted_in_disk.utils import read_iter_from_file

sid = sorted_in_disk(read_iter_from_file("path/to/file/to/read"),
                     key=lambda line: line.split(",")[1])

And to write sorted content in a different file:

from sorted_in_disk.utils import write_iter_in_file

count = write_iter_in_file("path/to/file/to/write", sid)

print("Total sorted lines: {}".format(count))

Limitations

Not mix mono-process and multiprocess configuration in same temporal file (only if you use only_one_read=False and append=True). Each mode (mono-process or multiprocess) work similar but save data in different ways to improve performance of each thread.

About this algorithm

This is an own algorithm create with my own experience in big data. This is invented by me and my experience, I do not investigated thirds or used any others to create this one. In addition, I use QuickQueue for Python than I invented and developed too.

Is useful for you?

Maybe you would be so kind to consider the amount of hours puts in, the great effort and the resources expended in doing this project. Thank you.

paypal

Contact

Contact me via: r.invarato@gmail.com

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

sorted-in-disk-1.0.3.tar.gz (31.5 kB view hashes)

Uploaded Source

Built Distribution

sorted_in_disk-1.0.3-py3-none-any.whl (23.5 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page