[
https://issues.apache.org/jira/browse/CASSANDRA-21745?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Nadav Har'El updated CASSANDRA-21745:
-------------------------------------
Description:
In CQL, a value of almost any type can be an "empty value" - a zero-length
value, which is different from null. For most non-string types (numbers, UUIDs,
dates, ...) this is a historical accident that makes no sense, and as
CASSANDRA-8354 explains, it is a nuisance. Users shouldn't be able to write
such values by accident - but currently, an empty string {{""}} written where a
value of a non-string type is expected is sometimes silently accepted as an
empty value, and sometimes rejected, depending on the type and on how the value
is written.
All the results below were tested on Cassandra 5.0.9 and 6.0-alpha2, with
identical results.
h3. INSERT JSON and fromJson()
INSERT JSON and the fromJson() function accept a JSON string for many
non-string types, e.g., "123" for an {{int}}. With a table having one column of
each type, using {{INSERT INTO t JSON '{"k": 1, "v": ""}'}} and {{INSERT INTO t
(k, v) VALUES (1, fromJson('""'))}}:
||Type||INSERT JSON with ""||fromJson('""')||
|int, bigint, boolean, decimal, double, float, uuid, timeuuid, inet,
timestamp|accepted (empty value)|accepted (empty value)|
|smallint, tinyint, duration|accepted (empty value)|server error
(RuntimeException, see below)|
|varint|InvalidRequest ("Value '' is not a valid representation of a varint
value")|FunctionFailure|
|blob, date, time|InvalidRequest|FunctionFailure|
The server errors for smallint, tinyint and duration happen because fromJson()
returns an empty value, which these types don't accept, so FunctionCall's check
of the return value throws a RuntimeException ("Return of function
system.fromjson : (text) -> smallint () is not a valid value for its declared
return type smallint") - an instance of CASSANDRA-18484.
h3. CQL string literals
The same inconsistency exists also for plain CQL string literals. The
non-string types date, time, inet and timestamp accept a string literal (e.g.,
'2024-01-01' for a date), but an empty string literal is rejected for date and
time, and accepted for inet and timestamp:
{code}
CREATE TABLE t (k int PRIMARY KEY, d date, t time, i inet, ts timestamp);
INSERT INTO t (k, d) VALUES (1, ''); -- InvalidRequest: Unable to coerce ''
to a formatted date (long)
INSERT INTO t (k, t) VALUES (1, ''); -- InvalidRequest: (TimeType) Unable to
coerce '' to a formatted time (long)
INSERT INTO t (k, i) VALUES (1, ''); -- accepted, writes an empty value
INSERT INTO t (k, ts) VALUES (1, ''); -- accepted, writes an empty value
{code}
h3. Suggested fix
An empty string is not a valid integer, float, UUID, boolean, IP address,
timestamp, etc. I suggest that an empty string be rejected as a value of every
type except the string types (ascii and text/varchar), consistently in INSERT
JSON, fromJson() and CQL string literals - as Cassandra already does for some
of these types (varint, blob, date, time). The user most likely made a mistake,
and almost certainly didn't intend to write an empty value. An empty value can
still be written deliberately, e.g., with {{blobAsInt(0x)}} or by binding an
empty value.
was:
In INSERT JSON and in the fromJson() function, a JSON string can be used as the
value of many non-string types - e.g., "123" for an {{int}} column. But the
*empty* string {{""}} is handled inconsistently between types: for some types
it is silently accepted and writes an "empty value" (a zero-length value,
different from null), and for other types it is rejected.
Tested on Cassandra 5.0.9 and 6.0-alpha2 (same behavior), with a table having
one column of each type, using
INSERT INTO t JSON '\{"k": 1, "v": ""}'
and
{{{}INSERT INTO t (k, v) VALUES (1, fromJson('""')){}}}:
||Type||INSERT JSON with ""||fromJson('""')||
|int, bigint, boolean, decimal, double, float, uuid, timeuuid|accepted (empty
value)|accepted (empty value)|
|smallint, tinyint, duration|accepted (empty value)|server error
(RuntimeException, see below)|
|varint|InvalidRequest ("Value '' is not a valid representation of a varint
value")|FunctionFailure|
|blob|InvalidRequest|FunctionFailure|
An empty string is not a valid integer, float, UUID or boolean, so I think it
should be rejected for all these types, as it already is for {{varint}} and
{{blob}} - the user most likely made a mistake, and almost certainly didn't
intend to write an empty value. As CASSANDRA-8354 explains, empty values of
these types exist only for historical reasons and are a nuisance; they
shouldn't be this easy to write by accident.
The server errors for smallint, tinyint and duration happen because fromJson()
returns an empty value, which these types don't accept, so FunctionCall's check
of the return value throws a RuntimeException ("Return of function
system.fromjson : (text) -> smallint () is not a valid value for its declared
return type smallint") - an instance of CASSANDRA-18484.
The same behavior for date, time, inet and timestamp (which also have string
literal forms) may deserve a separate discussion: on the tested versions, "" is
rejected for date and time but accepted for inet and timestamp.
> Empty string inconsistently accepted as a value of non-string types (INSERT
> JSON, fromJson(), string literals)
> --------------------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-21745
> URL: https://issues.apache.org/jira/browse/CASSANDRA-21745
> Project: Apache Cassandra
> Issue Type: Bug
> Reporter: Nadav Har'El
> Priority: Normal
>
> In CQL, a value of almost any type can be an "empty value" - a zero-length
> value, which is different from null. For most non-string types (numbers,
> UUIDs, dates, ...) this is a historical accident that makes no sense, and as
> CASSANDRA-8354 explains, it is a nuisance. Users shouldn't be able to write
> such values by accident - but currently, an empty string {{""}} written where
> a value of a non-string type is expected is sometimes silently accepted as an
> empty value, and sometimes rejected, depending on the type and on how the
> value is written.
> All the results below were tested on Cassandra 5.0.9 and 6.0-alpha2, with
> identical results.
> h3. INSERT JSON and fromJson()
> INSERT JSON and the fromJson() function accept a JSON string for many
> non-string types, e.g., "123" for an {{int}}. With a table having one column
> of each type, using {{INSERT INTO t JSON '{"k": 1, "v": ""}'}} and {{INSERT
> INTO t (k, v) VALUES (1, fromJson('""'))}}:
> ||Type||INSERT JSON with ""||fromJson('""')||
> |int, bigint, boolean, decimal, double, float, uuid, timeuuid, inet,
> timestamp|accepted (empty value)|accepted (empty value)|
> |smallint, tinyint, duration|accepted (empty value)|server error
> (RuntimeException, see below)|
> |varint|InvalidRequest ("Value '' is not a valid representation of a varint
> value")|FunctionFailure|
> |blob, date, time|InvalidRequest|FunctionFailure|
> The server errors for smallint, tinyint and duration happen because
> fromJson() returns an empty value, which these types don't accept, so
> FunctionCall's check of the return value throws a RuntimeException ("Return
> of function system.fromjson : (text) -> smallint () is not a valid value for
> its declared return type smallint") - an instance of CASSANDRA-18484.
> h3. CQL string literals
> The same inconsistency exists also for plain CQL string literals. The
> non-string types date, time, inet and timestamp accept a string literal
> (e.g., '2024-01-01' for a date), but an empty string literal is rejected for
> date and time, and accepted for inet and timestamp:
> {code}
> CREATE TABLE t (k int PRIMARY KEY, d date, t time, i inet, ts timestamp);
> INSERT INTO t (k, d) VALUES (1, ''); -- InvalidRequest: Unable to coerce ''
> to a formatted date (long)
> INSERT INTO t (k, t) VALUES (1, ''); -- InvalidRequest: (TimeType) Unable
> to coerce '' to a formatted time (long)
> INSERT INTO t (k, i) VALUES (1, ''); -- accepted, writes an empty value
> INSERT INTO t (k, ts) VALUES (1, ''); -- accepted, writes an empty value
> {code}
> h3. Suggested fix
> An empty string is not a valid integer, float, UUID, boolean, IP address,
> timestamp, etc. I suggest that an empty string be rejected as a value of
> every type except the string types (ascii and text/varchar), consistently in
> INSERT JSON, fromJson() and CQL string literals - as Cassandra already does
> for some of these types (varint, blob, date, time). The user most likely made
> a mistake, and almost certainly didn't intend to write an empty value. An
> empty value can still be written deliberately, e.g., with {{blobAsInt(0x)}}
> or by binding an empty value.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
