From 0ddf8c7f6fd4241e26016327ec85c5b958bdc32c Mon Sep 17 00:00:00 2001 From: Andrew Burkett Date: Thu, 21 Apr 2016 15:01:44 -0700 Subject: [PATCH] scrapy: init at 1.0.5 --- lib/maintainers.nix | 1 + pkgs/top-level/python-packages.nix | 60 ++++++++++++++++++++++++++++++ 2 files changed, 61 insertions(+) diff --git a/lib/maintainers.nix b/lib/maintainers.nix index 717d04e2031..1140362c16e 100644 --- a/lib/maintainers.nix +++ b/lib/maintainers.nix @@ -102,6 +102,7 @@ dmalikov = "Dmitry Malikov "; dochang = "Desmond O. Chang "; doublec = "Chris Double "; + drewkett = "Andrew Burkett "; ebzzry = "Rommel Martinez "; ederoyd46 = "Matthew Brown "; eduarrrd = "Eduard Bachmakov "; diff --git a/pkgs/top-level/python-packages.nix b/pkgs/top-level/python-packages.nix index 3d1e2026a82..77974dff752 100644 --- a/pkgs/top-level/python-packages.nix +++ b/pkgs/top-level/python-packages.nix @@ -26423,4 +26423,64 @@ in modules // { }; }; + w3lib = buildPythonPackage rec { + name = "w3lib-${version}"; + version = "1.14.2"; + + buildInputs = with self ; [ six pytest ]; + + src = pkgs.fetchurl { + url = "https://pypi.python.org/packages/source/w/w3lib/${name}.tar.gz"; + sha256 = "bd87eae62d208eef70869951abf05e96a8ee559714074a485168de4c5b190004"; + }; + + meta = { + description = "A library of web-related functions"; + homepage = "https://github.com/scrapy/w3lib"; + license = licenses.bsd3; + maintainers = with maintainers; [ drewkett ]; + }; + }; + + queuelib = buildPythonPackage rec { + name = "queuelib-${version}"; + version = "1.4.2"; + + src = pkgs.fetchurl { + url = "https://pypi.python.org/packages/source/q/queuelib/${name}.tar.gz"; + sha256 = "a6829918157ed433fafa87b0bb1e93e3e63c885270166db5884a02c34c86f914"; + }; + + buildInputs = with self ; [ pytest ]; + + meta = { + description = "A collection of persistent (disk-based) queues for Python"; + homepage = "https://github.com/scrapy/queuelib"; + license = licenses.bsd3; + maintainers = with maintainers; [ drewkett ]; + }; + }; + + scrapy = buildPythonPackage rec { + name = "Scrapy-${version}"; + version = "1.0.5"; + + disabled = isPy3k; + + buildInputs = with self ; [ pytest ]; + propagatedBuildInputs = with self ; [ six twisted w3lib lxml cssselect queuelib pyopenssl service-identity ]; + + src = pkgs.fetchurl { + url = "https://pypi.python.org/packages/source/S/Scrapy/${name}.tar.gz"; + sha256 = "0a51c785a310d65f6e70285a2da56d48ef7d049bd7fd60a08eef05c52328ca96"; + }; + + meta = { + description = "A fast high-level web crawling and web scraping framework, used to crawl websites and extract structured data from their pages"; + homepage = "http://scrapy.org/"; + license = licenses.bsd3; + maintainers = with maintainers; [ drewkett ]; + platforms = platforms.linux; + }; + }; }