[
https://issues.apache.org/jira/browse/HIVE-24693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17285449#comment-17285449
]
David Mollitor edited comment on HIVE-24693 at 2/16/21, 7:46 PM:
-----------------------------------------------------------------
In working on this ticket, I learned something interesting:
{code:java|title=Timestamp.java}
private static final DateTimeFormatter PARSE_FORMATTER = new
DateTimeFormatterBuilder()
// Date
.appendValue(YEAR, 1, 10,
SignStyle.NORMAL).appendLiteral('-').appendValue(MONTH_OF_YEAR, 1, 2,
SignStyle.NORMAL) ...
private static final DateTimeFormatter PRINT_FORMATTER = new
DateTimeFormatterBuilder()
// Date and Time Parts
.append(DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm:ss")) ...
{code}
When the *PARSING* code is built, it uses *YEAR*. However the *FORMATTER* code
is using *yyyy*. The equivalence should be:
{{ChornoField.YEAR}} = "uuuu"
{{ChronoField.YEAR_OF_ERA}} = "yyyy"
So, what I ran into in my work on skipping the timestamp parsing is that I
stumbled on the fact that Hive is reading YEAR but displaying YEAR_OF_ERA,
which are not the same things. YEAR has negative dates, YEAR_OF_ERA does not
usually have negative dates, for example a "negative date" in YEAR_OF_ERA would
be +2000 BCE whereas YEAR would be -2000. So, Hive is kinda whacky and out of
sync currently for negative years.
was (Author: belugabehr):
In working on this ticket, I learned something interesting:
{java|title=Timestamp.java}
private static final DateTimeFormatter PARSE_FORMATTER = new
DateTimeFormatterBuilder()
// Date
.appendValue(YEAR, 1, 10,
SignStyle.NORMAL).appendLiteral('-').appendValue(MONTH_OF_YEAR, 1, 2,
SignStyle.NORMAL) ...
private static final DateTimeFormatter PRINT_FORMATTER = new
DateTimeFormatterBuilder()
// Date and Time Parts
.append(DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm:ss")) ...
{code}
When the *PARSING* code is built, it uses *YEAR*. However the *FORMATTER* code
is using *yyyy*. The equivalence should be:
{{ChornoField.YEAR}} = "uuuu"
{{ChronoField.YEAR_OF_ERA}} = "yyyy"
So, what I ran into in my work on skipping the timestamp parsing is that I
stumbled on the fact that Hive is reading YEAR but displaying YEAR_OF_ERA,
which are not the same things. YEAR has negative dates, YEAR_OF_ERA does not
usually have negative dates, for example a "negative date" in YEAR_OF_ERA would
be +2000 BCE whereas YEAR would be -2000. So, Hive is kinda whacky and out of
sync currently for negative years.
> Parquet Timestamp Values Read/Write Very Slow
> ---------------------------------------------
>
> Key: HIVE-24693
> URL: https://issues.apache.org/jira/browse/HIVE-24693
> Project: Hive
> Issue Type: Improvement
> Reporter: David Mollitor
> Assignee: David Mollitor
> Priority: Critical
> Labels: pull-request-available
> Time Spent: 5.5h
> Remaining Estimate: 0h
>
> Parquet {{DataWriteableWriter}} relias on {{NanoTimeUtils}} to convert a
> timestamp object into a binary value. The way in which it does this,... it
> calls {{toString()}} on the timestamp object, and then parses the String.
> This particular timestamp do not carry a timezone, so the string is something
> like:
> {{2021-21-03 12:32:23.0000...}}
> The parse code tries to parse the string assuming there is a time zone, and
> if not, falls-back and applies the provided "default time zone". As was
> noted in [HIVE-24353], if something fails to parse, it is very expensive to
> try to parse again. So, for each timestamp in the Parquet file, it:
> * Builds a string from the time stamp
> * Parses it (throws an exception, parses again)
> There is no need to do this kind of string manipulations/parsing, it should
> just be using the epoch millis/seconds/time stored internal to the Timestamp
> object.
> {code:java}
> // Converts Timestamp to TimestampTZ.
> public static TimestampTZ convert(Timestamp ts, ZoneId defaultTimeZone) {
> return parse(ts.toString(), defaultTimeZone);
> }
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
