コード例 #1
0
def _rfc3339_to_datetime_string_split(string):
    # type: (str) -> Optional[datetime.datetime]
    """
    rfc3339_to_datetime variation which utilizes string.split approach.

    Returns a date time from a rfc3339 formatted timestamp.

    This doesn't do any complex testing and assumes the string is well formed and in UTC (e.g.
    uses Z at the end rather than a time offset).

    @param string: a date/time in rfc3339 format, e.g. 2015-08-03T09:12:43.143757463Z
    @rtype datetime.datetime
    """
    # split the string in to main time and fractional component
    parts = string.split(".")

    # it's possible that the time does not have a fractional component
    # e.g 2015-08-03T09:12:43Z, in this case 'parts' will only have a
    # single element that should end in Z.  Strip the Z if it exists
    # so we can use the same format string for processing the main
    # date+time regardless of whether the time has a fractional component.
    if parts[0].endswith("Z"):
        parts[0] = parts[0][:-1]

    # create a datetime object
    try:
        date_parts, time_parts = parts[0].split("T")
        date_parts = date_parts.split("-")  # type: ignore
        time_parts = time_parts.split(":")  # type: ignore

        dt = datetime.datetime(
            int(date_parts[0]),
            int(date_parts[1]),
            int(date_parts[2]),
            int(time_parts[0]),
            int(time_parts[1]),
            int(time_parts[2]),
        )
    except Exception:
        return None

    # now add the fractional part
    if len(parts) > 1:
        fractions = parts[1]
        # if we had a fractional component it should terminate in a Z
        if not fractions.endswith("Z"):
            # we don't handle non UTC timezones yet
            if any(c in fractions for c in "+-"):
                return None
            return dt

        # remove the Z and just process the fraction.
        fractions = fractions[:-1]
        to_micros = 6 - len(fractions)
        micro = int(int(fractions) * 10**to_micros)
        dt = dt.replace(microsecond=micro)

    return dt
コード例 #2
0
def _rfc3339_to_nanoseconds_since_epoch_regex(string):
    # type: (str) -> Optional[int]
    """
    rfc3339_to_nanoseconds_since_epoch variation which utilizes regex approach.
    """
    # split the string in to main time and fractional component
    parts = string.split(".")

    # it's possible that the time does not have a fractional component
    # e.g 2015-08-03T09:12:43Z, in this case 'parts' will only have a
    # single element that should end in Z.  Strip the Z if it exists
    # so we can use the same format string for processing the main
    # date+time regardless of whether the time has a fractional component.
    if parts[0].endswith("Z"):
        parts[0] = parts[0][:-1]

    try:
        dt = datetime.datetime(*list(
            map(int,
                RFC3339_STR_REGEX.match(parts[0]).groups()))  # type: ignore
                               )
    except Exception:
        return None

    nano_seconds = (calendar.timegm(
        (dt.year, dt.month, dt.day, dt.hour, dt.minute, dt.second)) *
                    1000000000)

    nanos = 0

    # now add the fractional part
    if len(parts) > 1:
        fractions = parts[1]
        # if the fractional part doesn't end in Z we likely have a
        # malformed time, so just return the current value
        if not fractions.endswith("Z"):
            # we don't handle non UTC timezones yet
            if any(c in fractions for c in "+-"):
                return None

            return nano_seconds

        # strip the final 'Z' and use the final number for processing
        fractions = fractions[:-1]
        to_nanos = 9 - len(fractions)
        nanos = int(int(fractions) * 10**to_nanos)

    return nano_seconds + nanos
コード例 #3
0
def _rfc3339_to_nanoseconds_since_epoch_strptime(string):
    # type: (str) -> Optional[int]
    """
    rfc3339_to_nanoseconds_since_epoch variation which utilizes strptime approach.
    """
    # split the string in to main time and fractional component
    parts = string.split(".")

    # it's possible that the time does not have a fractional component
    # e.g 2015-08-03T09:12:43Z, in this case 'parts' will only have a
    # single element that should end in Z.  Strip the Z if it exists
    # so we can use the same format string for processing the main
    # date+time regardless of whether the time has a fractional component.
    if parts[0].endswith("Z"):
        parts[0] = parts[0][:-1]

    try:
        tm = time.strptime(parts[0], "%Y-%m-%dT%H:%M:%S")
    except ValueError:
        return None

    nano_seconds = int(calendar.timegm(tm[0:6])) * 1000000000

    nanos = 0

    # now add the fractional part
    if len(parts) > 1:
        fractions = parts[1]
        # if the fractional part doesn't end in Z we likely have a
        # malformed time, so just return the current value
        if not fractions.endswith("Z"):
            # we don't handle non UTC timezones yet
            if any(c in fractions for c in "+-"):
                return None

            return nano_seconds

        # strip the final 'Z' and use the final number for processing
        fractions = fractions[:-1]
        to_nanos = 9 - len(fractions)
        nanos = int(int(fractions) * 10**to_nanos)

    return nano_seconds + nanos
コード例 #4
0
def _rfc3339_to_datetime_udatetime(string):
    # type: (str) -> Optional[datetime.datetime]
    """
    rfc3339_to_datetime variation which utilizes udatetime library.
    """
    # split the string in to main time and fractional component
    parts = string.split(".")

    # it's possible that the time does not have a fractional component
    # e.g 2015-08-03T09:12:43Z, in this case 'parts' will only have a
    # single element that should end in Z.  Strip the Z if it exists
    # so we can use the same format string for processing the main
    # date+time regardless of whether the time has a fractional component.
    if parts[0].endswith("Z"):
        parts[0] = parts[0][:-1]

    # create a datetime object
    try:
        dt = udatetime.from_string(parts[0])
        # NOTE: At this point we don't support timezones
        dt = dt.replace(tzinfo=None)
    except ValueError:
        return None

    # now add the fractional part
    if len(parts) > 1:
        fractions = parts[1]
        # if we had a fractional component it should terminate in a Z
        if not fractions.endswith("Z"):
            # we don't handle non UTC timezones yet
            if any(c in fractions for c in "+-"):
                return None
            return dt

        # remove the Z and just process the fraction.
        fractions = fractions[:-1]
        to_micros = 6 - len(fractions)
        micro = int(int(fractions) * 10**to_micros)
        # NOTE(Tomaz): dt.replace is quite slow...
        dt = dt.replace(microsecond=micro)

    return dt
コード例 #5
0
def _rfc3339_to_datetime_regex(string):
    # type: (str) -> Optional[datetime.datetime]
    """
    rfc3339_to_datetime variation which utilizes regex approach.
    """
    # split the string in to main time and fractional component
    parts = string.split(".")

    # it's possible that the time does not have a fractional component
    # e.g 2015-08-03T09:12:43Z, in this case 'parts' will only have a
    # single element that should end in Z.  Strip the Z if it exists
    # so we can use the same format string for processing the main
    # date+time regardless of whether the time has a fractional component.
    if parts[0].endswith("Z"):
        parts[0] = parts[0][:-1]

    # create a datetime object
    try:
        dt = datetime.datetime(*list(
            map(int,
                RFC3339_STR_REGEX.match(parts[0]).groups()))  # type: ignore
                               )
    except Exception:
        return None

    # now add the fractional part
    if len(parts) > 1:
        fractions = parts[1]
        # if we had a fractional component it should terminate in a Z
        if not fractions.endswith("Z"):
            # we don't handle non UTC timezones yet
            if any(c in fractions for c in "+-"):
                return None
            return dt

        # remove the Z and just process the fraction.
        fractions = fractions[:-1]
        to_micros = 6 - len(fractions)
        micro = int(int(fractions) * 10**to_micros)
        dt = dt.replace(microsecond=micro)

    return dt
コード例 #6
0
def _rfc3339_to_nanoseconds_since_epoch_string_split(string):
    # type: (str) -> Optional[int]
    """
    rfc3339_to_nanoseconds_since_epoch variation which utilizes string.split approach.

    Returns nanoseconds from unix epoch from a rfc3339 formatted timestamp.

    This doesn't do any complex testing and assumes the string is well formed and in UTC (e.g.
    uses Z at the end rather than a time offset).

    @param string: a date/time in rfc3339 format, e.g. 2015-08-03T09:12:43.143757463Z
    @rtype int
    """
    # split the string in to main time and fractional component
    parts = string.split(".")

    # it's possible that the time does not have a fractional component
    # e.g 2015-08-03T09:12:43Z, in this case 'parts' will only have a
    # single element that should end in Z.  Strip the Z if it exists
    # so we can use the same format string for processing the main
    # date+time regardless of whether the time has a fractional component.
    if parts[0].endswith("Z"):
        parts[0] = parts[0][:-1]

    try:
        date_parts, time_parts = parts[0].split("T")
        date_parts = date_parts.split("-")  # type: ignore
        time_parts = time_parts.split(":")  # type: ignore

        dt = datetime.datetime(
            int(date_parts[0]),
            int(date_parts[1]),
            int(date_parts[2]),
            int(time_parts[0]),
            int(time_parts[1]),
            int(time_parts[2]),
        )
    except Exception:
        return None

    nano_seconds = (calendar.timegm(
        (dt.year, dt.month, dt.day, dt.hour, dt.minute, dt.second)) *
                    1000000000)

    nanos = 0

    # now add the fractional part
    if len(parts) > 1:
        fractions = parts[1]
        # if the fractional part doesn't end in Z we likely have a
        # malformed time, so just return the current value
        if not fractions.endswith("Z"):
            # we don't handle non UTC timezones yet
            if any(c in fractions for c in "+-"):
                return None

            return nano_seconds

        # strip the final 'Z' and use the final number for processing
        fractions = fractions[:-1]
        to_nanos = 9 - len(fractions)
        nanos = int(int(fractions) * 10**to_nanos)

    return nano_seconds + nanos