First implementation of parallel loading, each thread loading a part of the input file (stat, seek, etc)

debian/changelog

@@ -1,3 +1,11 @@
+pgloader (2.3.0-1) unstable; urgency=low
+
+  * Mutli-threaded pgloader (see options max_threads, max_parallel_sections,
+                             section_threads and split_file_reading)
+  * FIX for -C and -I options (replace sys.log with self.log), per user request
+
+ -- Dimitri Fontaine <dim@tapoueh.org>  Mon, 11 Feb 2008 15:04:40 +0100
+
 pgloader (2.2.6-1) unstable; urgency=low
 
   * pgloader -V now VACUUM each table separately, no more vacuumdb issued

examples/parallel/parallel.sql

@@ -0,0 +1,11 @@
+CREATE TABLE parallel (
+ a integer primary key,
+ b text
+);
+
+-- create the .data file
+insert into parallel 
+       select * from (select a, a::text
+                        from generate_series(0, 1000 * 1000 * 1000) as t(a)) x;
+
+\copy parallel to 'parallel/parallel.data' with delimiter ';' csv

examples/pgloader.conf

@@ -12,13 +12,16 @@ lc_messages         = C
 
 ;client_encoding = 'utf-8'
 client_encoding = 'latin1'
-copy_every      = 5
+copy_every      = 5000
 commit_every    = 5
 #copy_delimiter  = %
 
 null         = ""
 empty_string = "\ "
 
+max_parallel_sections = 1
+
+
 [simple_tmpl]
 template     = True
 format       = text
@@ -96,6 +99,26 @@ field_sep       = |
 columns         = id, timestamp
 reformat        = timestamp:mysql:timestamp
 
+[parallel_template]
+template  = True
+table     = parallel
+format    = csv
+filename  = parallel/parallel.data
+field_sep = ;
+columns   = a, b
+
+[parallel_greg]
+use_template          = parallel_template
+max_threads           = 4
+section_threads       = -1
+split_file_reading    = True
+
+[parallel_simon]
+use_template          = parallel_template
+max_threads           = 2
+section_threads       = -1
+split_file_reading    = False
+
 [csv]
 table        = csv
 format       = csv

pgloader.py

@@ -273,6 +273,10 @@ def parse_config(conffile):
             rpath = config.get(section, 'reformat_path')
             pgloader.options.REFORMAT_PATH = rpath
 
+    if config.has_option(section, 'max_parallel_sections'):
+        mps = config.getint(section, 'max_parallel_sections')
+        pgloader.options.MAX_PARALLEL_SECTIONS = mps
+
     return config
 
 def myprint(l, line_prefix = "  ", cols = 78):
@@ -425,6 +429,7 @@ def load_data():
     # load some pgloader package modules
     from pgloader.options  import VERBOSE, DEBUG, QUIET, SUMMARY
     from pgloader.options  import DRY_RUN, PEDANTIC, VACUUM
+    from pgloader.options  import MAX_PARALLEL_SECTIONS
     from pgloader.pgloader import PGLoader
     from pgloader.tools    import PGLoader_Error
 
@@ -454,10 +459,32 @@ def load_data():
 
     threads = {}
     running = 0
-    for s in sections:
+    current = 0
-        summary[s] = []
+    
-        loader     = PGLoader(s, config, summary[s])
+    while current < len(sections):
+        s = sections[current]
+
+        # update running
+        if running > 0:
+            for s in threads:
+                if not threads[s].isAlive():
+                    running -= 1
+
+        if MAX_PARALLEL_SECTIONS != -1:
+            # -1 means we can start as many parallel section
+            # processing as we want to
+            
+            if running == MAX_PARALLEL_SECTIONS:
+                # we have to wait for one thread to terminate
+                # before considering next one
+                log.info('%d/%d threads running, sleeping %gs' \
+                         % (running, MAX_PARALLEL_SECTIONS, .1))
+                time.sleep(.1)
+                continue
+                        
         try:
+            summary[s] = []
+            loader     = PGLoader(s, config, summary[s])
             if not loader.template:
                 filename       = loader.filename
                 input_encoding = loader.input_encoding
@@ -489,17 +516,21 @@ def load_data():
         except KeyboardInterrupt:
             log.warning("Aborting on user demand (Interrupt)")
 
+        current += 1
+
     while running > 0:
         for s in threads:
             if not threads[s].isAlive():
                 running -= 1
 
         if running > 0:
-            log.info("%d thread(s) still running" % running)            
+            if MAX_PARALLEL_SECTIONS != 1:
+                log.info("%d thread(s) still running" % running)            
                 
             try:
-                log.info('waiting for %d threads, sleeping %gs' % (running, .1))
+                if MAX_PARALLEL_SECTIONS != 1:
-                time.sleep(.1)
+                    log.info('waiting for %d threads, sleeping %gs' % (running, 1))
+                time.sleep(1)
             except KeyboardInterrupt:
                 log.warning("Aborting %d threads still running at user demand"\
                             % running)

pgloader/csvreader.py

@@ -76,7 +76,17 @@ class CSVReader(DataReader):
             except IOError, error:
                 raise PGLoader_Error, error
 
+        if self.start is not None and self.start > 0:
+            self.log.info("CSV Reader starting at offset %d" % self.start)
+            fd.seek(self.start)
+
         # now read the lines
         for columns in csv.reader(fd, dialect = 'pgloader'):
+
+            if self.end is not None and fd.tell() >= self.end:
+                self.log.info("CSV Reader stoping, offset %d >= %d" % (fd.tell(), self.end()))
+                fd.close()
+                break
+            
             line = self.field_sep.join(columns)
             yield line, columns

pgloader/options.py

@@ -2,7 +2,7 @@
 #
 # Some common options, for each module to get them
 
-PGLOADER_VERSION = '2.2.6'
+PGLOADER_VERSION = '2.3.0~dev'
 
 INPUT_ENCODING     = None
 PG_CLIENT_ENCODING = 'latin9'
@@ -36,6 +36,11 @@ UDC_PREFIX = 'udc_'
 REFORMAT_PATH = None
 DEFAULT_REFORMAT_PATH = ['/usr/share/python-support/pgloader/reformat']
 
+MAX_THREADS           = 1
+MAX_PARALLEL_SECTIONS = 1
+SECTION_THREADS       = -1
+SPLIT_FILE_READING    = False
+
 CLIENT_MIN_MESSAGES = None
 LOG_MIN_MESSAGES    = DEBUG
 DEFAULT_LOG_FILE    = "/tmp/pgloader.log"

pgloader/pgloader.py

@@ -21,6 +21,8 @@ from options import COPY_SEP, FIELD_SEP, CLOB_SEP, NULL, EMPTY_STRING
 from options import NEWLINE_ESCAPES
 from options import UDC_PREFIX
 from options import REFORMAT_PATH
+from options import MAX_THREADS, MAX_PARALLEL_SECTIONS
+from options import SECTION_THREADS, SPLIT_FILE_READING
 
 class PGLoader(threading.Thread):
     """
@@ -28,14 +30,18 @@ class PGLoader(threading.Thread):
     import data with COPY or update blob data with UPDATE.
     """
 
-    def __init__(self, name, config, stats):
+    def __init__(self, name, config, stats, logname = None):
         """ Init with a configuration section """
         threading.Thread.__init__(self, name = name)
         
         # Some settings
         self.stats   = stats
         self.name    = name
-        self.log     = getLogger(name)
+        self.config  = config
+
+        if logname is None:
+            logname  = name
+        self.log     = getLogger(logname)
 
         self.__dbconnect__(config)
 
@@ -439,6 +445,42 @@ class PGLoader(threading.Thread):
             self.newline_escapes = [(a, NEWLINE_ESCAPES)
                                     for (a, x) in self.columns]
 
+        ##
+        # Parallelism knobs
+        for opt, default in [('max_threads', MAX_THREADS),
+                             ('section_threads', SECTION_THREADS),
+                             ('split_file_reading', SPLIT_FILE_READING)]:
+            
+            if config.has_option(name, opt):
+                if opt in ['max_threads', 'section_threads']:
+                    self.__dict__[opt] = config.getint(name, opt)
+                else:
+                    self.__dict__[opt] = config.get(name, opt) == 'True'
+            else:
+                if not self.template:
+                    self.__dict__[opt] = default
+
+            if not self.template:
+                self.log.info('%s.%s = %s' % (name, opt, str(self.__dict__[opt])))
+
+        if not self.template and self.split_file_reading:
+            global FROM_COUNT
+            if FROM_COUNT is not None and FROM_COUNT > 0:
+                raise PGLoader_Error, \
+                      "Conflict: can't use both 'split_file_reading' and '--from'"
+
+            global FROM_ID
+            if FROM_ID is not None:
+                raise PGLoader_Error, \
+                      "Conflict: can't use both 'split_file_reading' and '--from-id'"
+
+        if not self.template and self.section_threads > self.max_threads:
+            raise PGLoader_Error, \
+                  "%s.section_threads > %s.max_threads : %d > %d" \
+                  % (name, name, self.section_threads, self.max_threads)
+
+        ##
+        # Reader's init
         if config.has_option(name, 'format'):
             self.format = config.get(name, 'format')
 
@@ -648,11 +690,113 @@ class PGLoader(threading.Thread):
         return
 
     def run(self):
-        """ depending on configuration, do given job """
+        """ controling thread which dispatch the job """
         
         # Announce the beginning of the work
         self.log.info("[%s]" % self.name)
 
+        if self.max_threads == 1:
+
+            if self.reader.start is not None:
+                self.log.info("Loading from offset %d to %d" \
+                              %  (self.reader.start, self.reader.end))
+
+            self.prepare_processing()
+            self.process()
+            self.finish_processing()
+            return
+
+        # now we're going to need mutli-threading
+        if self.section_threads == -1:
+            self.section_threads = self.max_threads
+
+        if self.split_file_reading:
+            # this option is not compatible with text mode when
+            # field_count is used (meaning end of line could be found
+            # in the data)
+            if self.format.lower() == 'text' and self.field_count is not None:
+                raise PGLoader_Error, \
+                      "Can't use split_file_reading with text " +\
+                      "format when 'field_count' is used"
+            
+            # init boundaries to give to each thread
+            from stat import ST_SIZE
+            previous   = 0
+            filesize   = os.stat(self.filename)[ST_SIZE]
+            boundaries = []
+            for partn in range(self.section_threads):
+                start = previous
+                end   = (partn+1)*filesize / self.section_threads
+                boundaries.append((start, end))
+
+                previous = end + 1
+
+            self.log.info("Spliting input file of %d bytes %s" \
+                          % (filesize, str(boundaries)))
+
+            # Now check for real boundaries
+            fd = file(self.filename)
+            b  = 0
+            for b in range(len(boundaries)):
+                start, end = boundaries[b]
+                fd.seek(end)
+                dummy_str = fd.readline()
+
+                # update both current boundary end and next start
+                boundaries[b] = (start, fd.tell())
+                if (b+1) < len(boundaries):
+                    boundaries[b+1] = (fd.tell()+1, boundaries[b+1][1])
+                
+            fd.close()
+
+            self.log.info("Spliting input file of %d bytes %s" \
+                          % (filesize, str(boundaries)))
+
+            self.prepare_processing()
+
+            # now create self.section_threads PGLoader threads
+            summary = {}
+            threads = {}
+            running = 0
+            
+            for current in range(self.section_threads):
+                summary[current] = []
+                current_name     = "%s[%d]" % (self.name, current) 
+                loader = PGLoader(self.name,
+                                  self.config,
+                                  summary[current],
+                                  current_name)
+                loader.max_threads = 1
+                loader.reader.set_boundaries(boundaries[current])
+                loader.dont_prepare_nor_finish = True
+                
+                threads[current_name] = loader
+                threads[current_name].start()
+                running += 1
+
+            # wait for loaders completion
+            while running > 0:
+                for cn in threads:
+                    if not threads[cn].isAlive():
+                        running -= 1
+
+                if running > 0:
+                    log.info('waiting for %d threads, sleeping %gs' % (running, 1))
+                    time.sleep(1)
+
+            self.finish_processing()
+            log.info('No more threads are running, %s done' % self.name)
+            return
+
+        else:
+            # here we need a special thread reading the file
+            pass
+
+    def prepare_processing(self):
+        """ Things to do before processing data """
+        if 'dont_prepare_nor_finish' in self.__dict__:
+            return
+        
         if not DRY_RUN:
             if TRUNCATE:
                 self.db.truncate(self.table)
@@ -660,13 +804,10 @@ class PGLoader(threading.Thread):
             if TRIGGERS:
                 self.db.disable_triggers(self.table)
 
-        if self.columns is not None:
+    def finish_processing(self):
-            self.log.info("COPY csv data")
+        """ Things to do after processing data """
-            self.data_import()
+        if 'dont_prepare_nor_finish' in self.__dict__:
-
+            return
-        elif self.blob_cols is not None:
-            # elif: COPY process also blob data
-            self.log.info("UPDATE blob data")
 
         if TRIGGERS and not DRY_RUN:
             self.db.enable_triggers(self.table)
@@ -686,6 +827,18 @@ class PGLoader(threading.Thread):
         self.log.info("loading done")
         return
 
+
+    def process(self):
+        """ depending on configuration, do given job """
+
+        if self.columns is not None:
+            self.log.info("COPY csv data")
+            self.data_import()
+
+        elif self.blob_cols is not None:
+            # elif: COPY process also blob data
+            self.log.info("UPDATE blob data")
+
     def data_import(self):
         """ import CSV or TEXT data, using COPY """
 

pgloader/reader.py

@@ -37,6 +37,9 @@ class DataReader:
             if INPUT_ENCODING is not None:
                 self.input_encoding = INPUT_ENCODING
 
+        self.start = None
+        self.end   = None
+
     def readconfig(self, name, config):
         """ read configuration section for common options
 
@@ -84,3 +87,8 @@ class DataReader:
         """ read data from configured file, and generate (yields) for
         each data line: line, columns and rowid """
         pass
+
+    def set_boundaries(self, (start, end)):
+        """ set the boundaries of this reader """
+        self.start = start
+        self.end   = end

pgloader/textreader.py

@@ -96,10 +96,19 @@ class TextReader(DataReader):
             except IOError, error:
                 raise PGLoader_Error, error
 
+        if self.start is not None and self.start > 0:
+            self.log.info("Text Reader starting at offset %d" % self.start)
+            fd.seek(self.start)
+
         for line in fd:
             # we count real physical lines
             nb_plines += 1
 
+            if self.end is not None and fd.tell() >= self.end:
+                self.log.info("Text Reader stoping, offset %d >= %d" % (fd.tell(), self.end()))
+                fd.close()
+                break
+
             if self.input_encoding is not None:
                 # this may not be necessary, after all
                 try: