You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Andrey Mashenkov (Jira)" <ji...@apache.org> on 2021/05/19 13:03:00 UTC
[jira] [Created] (IGNITE-14743) Support large Tuples.
Andrey Mashenkov created IGNITE-14743:
-----------------------------------------
Summary: Support large Tuples.
Key: IGNITE-14743
URL: https://issues.apache.org/jira/browse/IGNITE-14743
Project: Ignite
Issue Type: Improvement
Reporter: Andrey Mashenkov
Fix For: 3.0
For now, TupleAssembler writes offsets for varlen columns as 2-byte {{short }}type.
This implicitly restricts key/value sizes down to 64 kB in total.
Let's use 4-bytes {{int}} type for offsets for large tuples.
Possible ways are:
# Just use ints for offsets, this increases the memory overhead for Rows.
# Pre-calculate potential row size during SchemaDescriptor initialization and keep 'offset_size' in schema.
Unlimited varlen type (which is default) usage will end-up user will have 4-byte offset size in most cases.
# Pre-calculate exact tuple size for each row and use row flags.
This requires Tuple data analysis which we already do to detect non-null varlen values and nulls. Strings may be a headache as we have to analyze each char for accurate tuple size calculation.
# Pre-calculate tuple size skipping chars analysis.
Using adaptive offset_size approaches allows us to use 1-2-4 byte numbers (byte, short, int) for offsets.
Collations for String columns may be introduced and used as a hint, but we will need to check a collation for every char on write.
'Large keys' is an unwanted case, may a solution for values only will be enough...
--
This message was sent by Atlassian Jira
(v8.3.4#803005)