This is an automated email from the ASF dual-hosted git repository.
Gabriel39 pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/doris-website.git
The following commit(s) were added to refs/heads/master by this push:
new 137a310cd3f [doc](lakehouse) Document timestamp compatibility changes
(#3982)
137a310cd3f is described below
commit 137a310cd3f1da52f1fccbce78a0b22508290922
Author: Gabriel <[email protected]>
AuthorDate: Thu Aug 6 12:10:07 2026 +0800
[doc](lakehouse) Document timestamp compatibility changes (#3982)
## What problem does this PR solve?
Related Doris PR: https://github.com/apache/doris/pull/65446
The Doris change updates timestamp handling in FileScannerV2 and Parquet
export. The user-visible compatibility rules and migration options must
be documented so users can distinguish raw Parquet `INT96` wall-clock
values from timestamps normalized by legacy writers.
## What is changed?
- Document that FileScannerV2 preserves raw Parquet `INT96` wall-clock
fields by default, while Parquet `INT64` logical timestamps retain their
standard semantics.
- Document the Hive Catalog and external-file TVF compatibility property
`hive.parquet.time-zone`, including validation, examples, `TIMESTAMPTZ`
handling, and its Hive-only scope; Iceberg is intentionally unaffected.
- Document that Hudi intentionally ignores `hive.parquet.time-zone` and
retains its existing SQL session-time-zone behavior for native base-file
and JNI merge-on-read scans.
- State that FileScannerV1 is intentionally unchanged and can differ
from FileScannerV2.
- Add an upgrade note for existing Hive tables whose legacy `INT96`
files were normalized by a known writer time zone; Hudi requires no
migration for this change.
- Document that Parquet Export and Outfile now default
`enable_int96_timestamps` to `false`, producing `INT64` logical
timestamps. Users needing Hive 2/3 compatibility can explicitly set it
to `true`.
- Document the explicit INT96 round-trip contract: when Doris exports
`INT96`, configure `hive.parquet.time-zone` with the export session time
zone when FileScannerV2 reads it back.
- Document round-half-up conversion from ORC nanoseconds to Doris
microseconds, including second carry and conservative predicate pruning.
- Keep the English and Simplified Chinese current documentation
synchronized.
Only current documentation is updated because apache/doris#65446 targets
unmerged master behavior. Published 4.x, 3.x, 2.1, and Japanese
documentation are intentionally unchanged.
## Validation
- `git diff --check` — passed.
- `yarn docs:features:changed` — passed with no findings.
- `yarn docs:i18n-sync:changed` — English/Chinese current pages are
synchronized; expected warnings remain for intentionally unchanged
published versions and Japanese docs.
- `yarn docs:links:changed` / `yarn docs:lint:changed` — no new link
issues from this change; the commands report existing missing Hive
Catalog anchors and existing repository-wide SEO/Markdown findings.
- `yarn key-features:generate`
- `./local_dev.sh build-all --versions current --skip-install --max-mem
8192` — generated English and Chinese current site output; the build
reports existing repository-wide broken-link warnings.
---
docs/data-operate/export/export-overview.md | 2 +-
docs/lakehouse/catalogs/hive-catalog.mdx | 46 ++++++++++++++++++++++
docs/lakehouse/file-formats/orc.md | 12 ++++++
docs/lakehouse/file-formats/parquet.md | 9 ++++-
.../current/data-operate/export/export-overview.md | 2 +-
.../current/lakehouse/catalogs/hive-catalog.mdx | 46 ++++++++++++++++++++++
.../current/lakehouse/file-formats/orc.md | 12 ++++++
.../current/lakehouse/file-formats/parquet.md | 9 ++++-
8 files changed, 134 insertions(+), 4 deletions(-)
diff --git a/docs/data-operate/export/export-overview.md
b/docs/data-operate/export/export-overview.md
index cadfecb956d..670da7fbf0c 100644
--- a/docs/data-operate/export/export-overview.md
+++ b/docs/data-operate/export/export-overview.md
@@ -143,7 +143,7 @@ When Apache Doris exports to the Parquet file format, it
first converts the in-m
| hll | binary | BYTE_ARRAY |
|
:::note
-In versions 2.1.11 and 3.0.7, the `parquet.enable_int96_timestamps` property
is supported to specify whether the Doris `datetimev2` type is stored as
`INT96` or `INT64` in Parquet. The default is `INT96`. `INT96` has been
deprecated in the Parquet standard and is used only for compatibility with
legacy systems (such as Hive versions before 4.0).
+For Parquet Export and Outfile, use the `enable_int96_timestamps` property to
choose the physical encoding of Doris `DATETIMEV2` values. The default is
`false`, which writes `INT64` with a Parquet timestamp logical type. Set it to
`true` to write the deprecated `INT96` encoding only when a legacy downstream
reader, such as Hive 2 or Hive 3, requires it. Asynchronous Export stores this
property with the export job and applies it when the job executes. If Doris
later reads an explicitly ex [...]
:::
## Related Documents
diff --git a/docs/lakehouse/catalogs/hive-catalog.mdx
b/docs/lakehouse/catalogs/hive-catalog.mdx
index 77f4772f437..47308948769 100644
--- a/docs/lakehouse/catalogs/hive-catalog.mdx
+++ b/docs/lakehouse/catalogs/hive-catalog.mdx
@@ -72,6 +72,8 @@ CREATE CATALOG [IF NOT EXISTS] catalog_name PROPERTIES (
* `hive.staging_dir`: Used to set the temporary staging root directory for
Hive write operations. During the write process, data is first written to this
directory and then moved to the final Hive table directory after completion.
This path can be either a relative path relative to the Hive table root
directory or an absolute path. The default value is `/tmp/.doris_staging`. In
some cases, when writing to Hive tables, you may encounter issues where the
`/tmp` directory and the table di [...]
+ * `hive.parquet.time-zone`: The compatibility time zone used to restore
legacy Parquet `INT96` timestamps when FileScannerV2 reads Hive tables. The
default is empty, which preserves the wall-clock fields stored in `INT96`
without applying a time zone conversion. Set this property only when the
`INT96` files in the Catalog were normalized by the writer with a known time
zone. Hudi tables intentionally ignore this property and retain their existing
SQL session-time-zone behavior. For det [...]
+
* `{CommonProperties}`
The CommonProperties section is for entering common attributes. Please see
the "Common Properties" section in the [Catalog
Overview](../catalog-overview.md).
@@ -208,6 +210,50 @@ Hive transactional tables are supported from version 3.x
onwards. For details, r
* [ORC](../file-formats/orc.md)
+## Parquet INT96 Timestamp Compatibility {#timestamp-compatibility}
+
+Hive tables can contain timestamp files produced by different engines. The
physical encoding and writer behavior determine whether a time zone conversion
is required. Doris does not infer a writer time zone from a Parquet `INT96`
value because `INT96` has no time zone annotation.
+
+### Parquet INT96 timestamps
+
+FileScannerV2 handles Parquet timestamps as follows:
+
+| Physical encoding | Default behavior | Effect of `hive.parquet.time-zone` |
+|---|---|---|
+| `INT96` mapped to `DATETIMEV2` | Preserve the date and time fields stored in
the file. The SQL session time zone does not shift the value. | Treat the raw
`INT96` value as a UTC instant and convert it to the configured compatibility
time zone, restoring timestamps that a legacy writer normalized. |
+| `INT64` with a timestamp logical type | Follow the Parquet logical-type
semantics. | Not affected. |
+
+For example, suppose the intended local timestamp is `2021-01-01 10:11:00` in
`Asia/Shanghai`:
+
+* If the file stores the wall-clock fields directly as `10:11:00`, leave
`hive.parquet.time-zone` unset. Doris returns `10:11:00` even when the SQL
session uses `Asia/Shanghai`.
+* If a legacy writer normalizes the value to UTC before storing `INT96`, the
file may contain `02:11:00`. Set `hive.parquet.time-zone` to `Asia/Shanghai` so
that Doris restores `10:11:00`.
+
+Configure the compatibility time zone when creating the Catalog:
+
+```sql
+CREATE CATALOG hive_legacy_int96 PROPERTIES (
+ 'type' = 'hms',
+ 'hive.metastore.uris' = 'thrift://127.0.0.1:9083',
+ 'hive.parquet.time-zone' = 'Asia/Shanghai'
+);
+```
+
+The value can be an IANA time zone name, such as `Asia/Shanghai`, or a UTC
offset from `-12:00` through `+14:00`. An invalid value causes Catalog creation
to fail.
+
+:::caution
+`hive.parquet.time-zone` applies to every `INT96` timestamp column mapped to
`DATETIMEV2` in Hive tables in the Catalog. It does not affect Iceberg tables
in the same Hive Metastore because Iceberg defines its own timestamp semantics.
Hudi also intentionally ignores this property: both native base-file scans and
JNI merge-on-read scans retain the pre-existing behavior of interpreting their
timestamps with the SQL session time zone. Parquet metadata cannot reliably
tell Doris which conven [...]
+:::
+
+The same property is accepted by file table-valued functions that read Parquet
files. The compatibility conversion applies only to Parquet `INT96` decoding in
FileScannerV2. FileScannerV1 is intentionally unchanged and can return
different results; switching to FileScannerV1 is not a compatibility mechanism
for this change. Parquet `INT64` logical timestamps are not affected. When an
`INT96` column maps to `TIMESTAMPTZ`, Doris ignores this property and preserves
the UTC instant; displayi [...]
+
+### Migration recommendations
+
+* Doris writes Parquet `INT64` logical timestamps by default. Set
`enable_int96_timestamps = true` only when a legacy reader such as Hive 2 or
Hive 3 requires `INT96`.
+* When upgrading to a version with this FileScannerV2 behavior, check existing
Hive tables before rollout. A Catalog containing `INT96` files that were
normalized by a legacy writer can return a different wall-clock value after the
upgrade. Compare representative values and set `hive.parquet.time-zone` to the
writer time zone when conversion is required. Hudi requires no migration for
this change because it retains its existing session-time-zone behavior.
+* Before setting `hive.parquet.time-zone`, inspect representative files and
compare known timestamp values. An unnecessary setting introduces the same kind
of offset that it is intended to correct.
+* Keep files with different `INT96` writer conventions in separate Catalogs,
or rewrite them to a consistent `INT64` encoding.
+* When Doris explicitly exports `INT96` with `enable_int96_timestamps = true`,
it normalizes timestamps using the export session time zone. To read the file
back through FileScannerV2, set `hive.parquet.time-zone` to that export session
time zone. See [Export Overview](../../data-operate/export/export-overview.md).
+
## Column Type Mapping
| Hive Type | Doris Type | Comment
|
diff --git a/docs/lakehouse/file-formats/orc.md
b/docs/lakehouse/file-formats/orc.md
index 63a8b3e8e0b..b64e01508dc 100644
--- a/docs/lakehouse/file-formats/orc.md
+++ b/docs/lakehouse/file-formats/orc.md
@@ -17,6 +17,18 @@ This document introduces the support for reading and writing
ORC file formats in
* Writing data during Export.
* Writing data with Outfile.
+## Timestamp Fractional-Second Precision
+
+ORC timestamps can store nanoseconds, while Doris `DATETIMEV2` and
`TIMESTAMPTZ` support up to microseconds. FileScannerV2 rounds ORC fractional
seconds to the nearest microsecond using round half up in Catalog scans,
table-valued functions, and Broker Load:
+
+| ORC value | Doris value |
+|---|---|
+| `2024-01-01 00:00:00.123456499` | `2024-01-01 00:00:00.123456` |
+| `2024-01-01 00:00:00.123456500` | `2024-01-01 00:00:00.123457` |
+| `2024-01-01 00:00:00.999999500` | `2024-01-01 00:00:01.000000` |
+
+Both `DATETIMEV2` and `TIMESTAMPTZ` mappings use this rule, including a carry
into the next second. Row decoding, statistics conversion, and predicate
pushdown use matching boundaries. When ORC statistics do not have enough
precision to represent a rounded boundary exactly, Doris widens the pruning
boundary conservatively. If a timestamp predicate such as `!=` cannot be
represented safely, Doris skips that search argument and evaluates the
predicate on decoded rows. This can reduce pruni [...]
+
## Supported Compression Formats
* uncompressed
diff --git a/docs/lakehouse/file-formats/parquet.md
b/docs/lakehouse/file-formats/parquet.md
index 48dc54ba429..55183f971a2 100644
--- a/docs/lakehouse/file-formats/parquet.md
+++ b/docs/lakehouse/file-formats/parquet.md
@@ -17,6 +17,14 @@ This document introduces the support for reading and writing
Parquet file format
* Writing data during Export.
* Writing data with Outfile.
+## INT96 Timestamp Decoding
+
+Parquet `INT96` stores date and time fields without a time zone annotation.
FileScannerV2 therefore preserves the raw wall-clock value by default instead
of shifting it with the SQL session time zone. For example, a raw `2021-01-01
10:11:00` remains `10:11:00` when mapped to `DATETIMEV2`, including in Catalog
scans, table-valued functions, and Broker Load.
+
+This behavior applies only to `INT96`. Parquet `INT64` values with a timestamp
logical type continue to follow the logical-type semantics. If legacy Hive
writers normalized `INT96` values with a known time zone, configure
`hive.parquet.time-zone` in the [Hive
Catalog](../catalogs/hive-catalog.mdx#timestamp-compatibility). External file
table-valued functions also accept that property. Other FileScannerV2 entry
points preserve the raw `INT96` wall-clock value, except Hudi, which intention
[...]
+
+When an `INT96` column maps to `TIMESTAMPTZ`, Doris preserves the UTC instant
rather than applying the compatibility time zone.
+
## Supported Compression Formats
* uncompressed
@@ -57,4 +65,3 @@ This document introduces the support for reading and writing
Parquet file format
The maximum buffer size allocated when reading a Column within a Parquet
Row Group. Default is 8M.
-
diff --git
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/data-operate/export/export-overview.md
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/data-operate/export/export-overview.md
index cc75b874bfb..31f5056f5aa 100644
---
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/data-operate/export/export-overview.md
+++
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/data-operate/export/export-overview.md
@@ -143,7 +143,7 @@ Apache Doris 导出到 Parquet 文件格式时，会先将 Doris 内存数据转
| hll | binary | BYTE_ARRAY |
|
:::note
-在 2.1.11 和 3.0.7 版本中，支持通过 `parquet.enable_int96_timestamps` 属性指定 Doris 的
`datetimev2` 类型在 Parquet 中是使用 `INT96` 还是 `INT64` 存储，默认为 `INT96`。`INT96` 在
Parquet 标准中已被废弃，仅用于兼容旧系统（如 Hive 4.0 之前的版本）。
+Parquet Export 和 Outfile 使用 `enable_int96_timestamps` 属性选择 Doris `DATETIMEV2`
值的物理编码。默认值为 `false`，写入带 Parquet 时间戳逻辑类型的 `INT64`。仅当 Hive 2、Hive 3
等下游旧版读取端要求时，才设置为 `true` 写入已废弃的 `INT96` 编码。异步 Export 会将该属性保存在导出作业中，并在作业执行时应用。如果
Doris 后续读取显式导出的 `INT96` 文件，请在 Hive Catalog 或文件表值函数中将 `hive.parquet.time-zone`
设置为导出会话时区，使 FileScannerV2 还原原始墙上时间。
:::
## 相关文档
diff --git
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/catalogs/hive-catalog.mdx
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/catalogs/hive-catalog.mdx
index 3220ff68b5a..04baee6e5b3 100644
---
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/catalogs/hive-catalog.mdx
+++
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/catalogs/hive-catalog.mdx
@@ -74,6 +74,8 @@ CREATE CATALOG [IF NOT EXISTS] catalog_name PROPERTIES (
* `hive.staging_dir`：用于设置 Hive 写操作的临时 staging
根目录。写入过程中数据会先写入该目录，写入完成后再移动到最终的 Hive 表目录。该路径可以是相对于 Hive 表根目录的相对路径，也可以是绝对路径。默认值为
`/tmp/.doris_staging`。某些情况下，写入 Hive 表时会遇到 `/tmp` 目录和表目录分属不同
namespace，导致无法写入的问题，此时可以通过修改此参数，使用相对路径解决该问题。该参数自 4.0.3 版本开始支持。
+ * `hive.parquet.time-zone`：FileScannerV2 读取 Hive 表时，用于还原旧版 Parquet `INT96`
时间戳的兼容时区。默认值为空，此时 Doris 保留 `INT96` 中存储的日期和时间字段，不做时区转换。仅当 Catalog 中的 `INT96`
文件由写入端按已知时区进行过归一化时，才需要设置此属性。Hudi 表会有意忽略该属性，并保留现有的 SQL 会话时区行为。详情请参阅 [Parquet
INT96 时间戳兼容性](#timestamp-compatibility)。
+
* `{CommonProperties}`
CommonProperties 部分用于填写通用属性。请参阅[ 数据目录概述 ](../catalog-overview.md)中【通用属性】部分。
@@ -224,6 +226,50 @@ ORDER BY entry_name;
* [ORC](../file-formats/orc.md)
+## Parquet INT96 时间戳兼容性 {#timestamp-compatibility}
+
+Hive 表中的时间戳文件可能由不同引擎生成。是否需要时区转换取决于物理编码和写入端行为。Parquet `INT96` 本身不包含时区标注，因此
Doris 不会根据 `INT96` 值推断写入端时区。
+
+### Parquet INT96 时间戳
+
+FileScannerV2 按以下规则处理 Parquet 时间戳：
+
+| 物理编码 | 默认行为 | `hive.parquet.time-zone` 的作用 |
+|---|---|---|
+| 映射到 `DATETIMEV2` 的 `INT96` | 保留文件中存储的日期和时间字段，SQL 会话时区不会使该值发生偏移。 | 将原始
`INT96` 值视为 UTC 时刻，再转换到配置的兼容时区，从而还原被旧版写入端归一化过的时间戳。 |
+| 带时间戳逻辑类型的 `INT64` | 遵循 Parquet 逻辑类型语义。 | 不受影响。 |
+
+例如，假设预期的本地时间戳是 `Asia/Shanghai` 时区的 `2021-01-01 10:11:00`：
+
+* 如果文件直接存储墙上时间字段 `10:11:00`，请不要设置 `hive.parquet.time-zone`。即使 SQL 会话使用
`Asia/Shanghai`，Doris 仍返回 `10:11:00`。
+* 如果旧版写入端先将该值归一化为 UTC 再写入 `INT96`，文件中可能存储 `02:11:00`。将
`hive.parquet.time-zone` 设置为 `Asia/Shanghai` 后，Doris 会还原出 `10:11:00`。
+
+创建 Catalog 时配置兼容时区：
+
+```sql
+CREATE CATALOG hive_legacy_int96 PROPERTIES (
+ 'type' = 'hms',
+ 'hive.metastore.uris' = 'thrift://127.0.0.1:9083',
+ 'hive.parquet.time-zone' = 'Asia/Shanghai'
+);
+```
+
+该值可以是 `Asia/Shanghai` 等 IANA 时区名称，也可以是 `-12:00` 到 `+14:00` 范围内的 UTC 偏移。非法值会导致
Catalog 创建失败。
+
+:::caution
+`hive.parquet.time-zone` 作用于该 Catalog 中所有 Hive 表内映射到 `DATETIMEV2` 的 `INT96`
时间戳列。该属性不影响同一 Hive Metastore 中的 Iceberg 表，因为 Iceberg 定义了自己的时间戳语义。Hudi
也会有意忽略该属性：原生 base file 扫描和 JNI merge-on-read 扫描都保留变更前的行为，使用 SQL
会话时区解释时间戳。Parquet 元数据无法可靠地告诉 Doris 某个 `INT96` 写入端使用了哪种约定。不要仅为了匹配 SQL
会话时区而设置此属性。如果同一个 Hive Metastore 中的 Hive 表包含使用不同约定写入的 `INT96` 文件，应尽量通过不同 Catalog
分别访问。
+:::
+
+读取 Parquet 文件的文件表值函数也接受同名属性。该兼容转换仅影响 FileScannerV2 对 Parquet `INT96`
的解码。FileScannerV1 有意保持不变，可能返回不同结果；切换到 FileScannerV1 不是本次变更的兼容方案。Parquet `INT64`
逻辑时间戳不受影响。`INT96` 列映射到 `TIMESTAMPTZ` 时，Doris 会忽略该属性并保留 UTC 时刻；该值的显示仍遵循
`TIMESTAMPTZ` 和会话时区语义。
+
+### 迁移建议
+
+* Doris 默认写入带逻辑类型的 Parquet `INT64` 时间戳。仅当 Hive 2、Hive 3 等旧版读取端要求 `INT96` 时，才设置
`enable_int96_timestamps = true`。
+* 升级到采用此 FileScannerV2 行为的版本前，请先检查现有 Hive 表。包含由旧版写入端归一化过的 `INT96`
文件时，升级后返回的墙上时间可能发生变化。请比对有代表性的已知值，并在需要转换时将 `hive.parquet.time-zone`
设置为写入端时区。Hudi 保留现有的会话时区行为，因此不需要针对本次变更执行迁移。
+* 设置 `hive.parquet.time-zone` 前，请检查有代表性的文件并与已知时间戳值进行比对。不必要的配置反而会产生原本希望消除的时间偏移。
+* 对使用不同 `INT96` 写入约定的文件，使用不同 Catalog 分别访问，或将文件重写为统一的 `INT64` 编码。
+* Doris 通过 `enable_int96_timestamps = true` 显式导出 `INT96` 时，会使用导出会话时区归一化时间戳。通过
FileScannerV2 读回该文件时，请将 `hive.parquet.time-zone`
设置为相同的导出会话时区。详情请参阅[数据导出概述](../../data-operate/export/export-overview.md)。
+
## 列类型映射
| Hive Type | Doris Type | Comment |
diff --git
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/orc.md
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/orc.md
index 7e2bfa5b36c..91230cc017a 100644
---
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/orc.md
+++
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/orc.md
@@ -21,6 +21,18 @@
* Outfile 中对数据的写入操作。
+## 时间戳小数秒精度
+
+ORC 时间戳可以保存纳秒，而 Doris `DATETIMEV2` 和 `TIMESTAMPTZ` 最高支持微秒。FileScannerV2 在
Catalog 扫描、表值函数和 Broker Load 中使用四舍五入规则，将 ORC 小数秒舍入到最接近的微秒：
+
+| ORC 值 | Doris 值 |
+|---|---|
+| `2024-01-01 00:00:00.123456499` | `2024-01-01 00:00:00.123456` |
+| `2024-01-01 00:00:00.123456500` | `2024-01-01 00:00:00.123457` |
+| `2024-01-01 00:00:00.999999500` | `2024-01-01 00:00:01.000000` |
+
+`DATETIMEV2` 和 `TIMESTAMPTZ` 映射均使用该规则，包括向下一秒进位。行解码、统计信息转换和谓词下推使用一致的边界。当 ORC
统计信息的精度不足以准确表示舍入边界时，Doris 会保守地扩大裁剪边界。如果 `!=` 等时间戳谓词无法安全表示，Doris
会跳过对应的搜索条件，并在解码后的数据行上计算谓词。这可能降低该谓词的裁剪效果，但可以避免错误跳过有效数据。
+
## 支持的压缩格式
* umcomressed
diff --git
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/parquet.md
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/parquet.md
index 2c029d1c5cf..b23661360d4 100644
---
a/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/parquet.md
+++
b/i18n/zh-CN/docusaurus-plugin-content-docs/current/lakehouse/file-formats/parquet.md
@@ -21,6 +21,14 @@
* Outfile 中对数据的写入操作。
+## INT96 时间戳解码
+
+Parquet `INT96` 存储日期和时间字段，但不包含时区标注。因此 FileScannerV2 默认保留原始墙上时间值，而不会根据 SQL
会话时区产生偏移。例如，映射到 `DATETIMEV2` 时，原始值 `2021-01-01 10:11:00` 在 Catalog 扫描、表值函数和
Broker Load 中均保持为 `10:11:00`。
+
+该行为仅适用于 `INT96`。带时间戳逻辑类型的 Parquet `INT64` 值仍遵循逻辑类型语义。如果旧版 Hive 写入端按已知时区归一化过
`INT96` 值，请在 [Hive
Catalog](../catalogs/hive-catalog.mdx#timestamp-compatibility) 中配置
`hive.parquet.time-zone`。外部文件表值函数也接受该属性。其他 FileScannerV2 入口会保留原始 `INT96`
墙上时间值，但 Hudi 除外：Hudi 的原生扫描和 JNI 扫描会有意保留原有的 SQL 会话时区行为，并忽略
`hive.parquet.time-zone`。
+
+`INT96` 列映射到 `TIMESTAMPTZ` 时，Doris 会保留 UTC 时刻，而不应用兼容时区。
+
## 支持的压缩格式
* umcomressed
@@ -66,4 +74,3 @@
* `parquet_column_max_buffer_mb` (2.1+, 3.0+)
读取 Parquet Row Group 中的 Column 时所分配的最大 Buffer 大小，默认为 8M。
-
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
