DoltLite keys user-table rows in its prolly tree by the user's primary key
columns rather than by SQLite's auto-allocated rowid (since #358). Rowid is
allocation-order dependent and therefore history dependent, which is a hard
incompatibility with the Dolt model: two branches that independently insert
identical rows have to produce identical row identities, and a
content-addressed tree cannot do that if the key is "whichever integer this
connection handed out first".
Consequences, all of which inherited SQLite testfixture suites exercise:
- a table without an explicit INTEGER PRIMARY KEY is auto-converted to
WITHOUT ROWID, so rowid is not an addressable column — WHERE rowid = N
and ORDER BY rowid error with "no such column: rowid"
- an unordered
SELECT returns user-PK order where stock SQLite returns
rowid (insertion) order
PRAGMA table_info reports non-PK columns differently for auto-converted
WITHOUT ROWID tables- AUTOINCREMENT and
sqlite_sequence behavior tied to rowid allocation
- foreign key
SET NULL / SET DEFAULT whose triggering DELETE addresses
parent rows by rowid
This is the single largest rationale in
test/known_testfixture_divergences.txt — its header block has documented it
at length, but with no issue to point at, so per-suite rows in
test/known_testfixture_exception_inventory.txt had nowhere to attribute it.
This issue is that anchor.
Not expected to be fixed; rowid-keyed storage is incompatible with the
version-control model rather than merely unimplemented.
Attribution status
Attributed so far are the suites whose own recorded reason names rowid
explicitly (25 divergence suites plus btreefault, incrblob, walbig).
That is a floor, not the true count: most rowid-driven suites rely on the
header block instead of repeating the reason inline. Widening the attribution
is a triage pass over the remaining intentional rows and should not be
guessed at in bulk.
DoltLite keys user-table rows in its prolly tree by the user's primary key
columns rather than by SQLite's auto-allocated rowid (since #358). Rowid is
allocation-order dependent and therefore history dependent, which is a hard
incompatibility with the Dolt model: two branches that independently insert
identical rows have to produce identical row identities, and a
content-addressed tree cannot do that if the key is "whichever integer this
connection handed out first".
Consequences, all of which inherited SQLite testfixture suites exercise:
WITHOUT ROWID, so
rowidis not an addressable column —WHERE rowid = Nand
ORDER BY rowiderror with "no such column: rowid"SELECTreturns user-PK order where stock SQLite returnsrowid (insertion) order
PRAGMA table_inforeports non-PK columns differently for auto-convertedWITHOUT ROWID tables
sqlite_sequencebehavior tied to rowid allocationSET NULL/SET DEFAULTwhose triggering DELETE addressesparent rows by rowid
This is the single largest rationale in
test/known_testfixture_divergences.txt— its header block has documented itat length, but with no issue to point at, so per-suite rows in
test/known_testfixture_exception_inventory.txthad nowhere to attribute it.This issue is that anchor.
Not expected to be fixed; rowid-keyed storage is incompatible with the
version-control model rather than merely unimplemented.
Attribution status
Attributed so far are the suites whose own recorded reason names rowid
explicitly (25 divergence suites plus
btreefault,incrblob,walbig).That is a floor, not the true count: most rowid-driven suites rely on the
header block instead of repeating the reason inline. Widening the attribution
is a triage pass over the remaining
intentionalrows and should not beguessed at in bulk.