periods.c

#include "postgres.h"
#include "fmgr.h"

#include "access/htup_details.h"
#include "access/heapam.h"
#if (PG_VERSION_NUM < 120000)
#define table_open(r, l)	heap_open(r, l)
#define table_close(r, l)	heap_close(r, l)
#else
#include "access/table.h"
#endif
#include "access/tupconvert.h"
#include "access/xact.h"
#include "catalog/pg_type.h"
#include "commands/trigger.h"
#include "datatype/timestamp.h"
#include "executor/spi.h"
#include "funcapi.h"
#include "lib/stringinfo.h"
#include "nodes/bitmapset.h"
#include "utils/builtins.h"
#include "utils/date.h"
#include "utils/datum.h"
#include "utils/elog.h"
#if (PG_VERSION_NUM < 100000)
#else
#include "utils/fmgrprotos.h"
#endif
#include "utils/hsearch.h"
#include "utils/lsyscache.h"
#include "utils/memutils.h"
#include "utils/rel.h"
#include "utils/timestamp.h"

PG_MODULE_MAGIC;

PGDLLEXPORT Datum generated_always_as_row_start_end(PG_FUNCTION_ARGS);
PGDLLEXPORT Datum write_history(PG_FUNCTION_ARGS);

PG_FUNCTION_INFO_V1(generated_always_as_row_start_end);
PG_FUNCTION_INFO_V1(write_history);

/* Define some SQLSTATEs that might not exist */
#if (PG_VERSION_NUM < 100000)
#define ERRCODE_GENERATED_ALWAYS MAKE_SQLSTATE('4','2','8','C','9')
#endif
#define ERRCODE_INVALID_ROW_VERSION MAKE_SQLSTATE('2','2','0','1','H')

/* We use these a lot, so make aliases for them */
#if (PG_VERSION_NUM < 100000)
#define TRANSACTION_TSTZ	TimestampTzGetDatum(GetCurrentTransactionStartTimestamp())
#define TRANSACTION_TS		DirectFunctionCall1(timestamptz_timestamp, TRANSACTION_TSTZ)
#define TRANSACTION_DATE	DirectFunctionCall1(timestamptz_date, TRANSACTION_TSTZ)
#else
#define TRANSACTION_TSTZ	TimestampTzGetDatum(GetCurrentTransactionStartTimestamp())
#define TRANSACTION_TS		DirectFunctionCall1(timestamptz_timestamp, TRANSACTION_TSTZ)
#define TRANSACTION_DATE	DateADTGetDatum(GetSQLCurrentDate())
#endif

#define INFINITE_TSTZ		TimestampTzGetDatum(DT_NOEND)
#define INFINITE_TS			TimestampGetDatum(DT_NOEND)
#define INFINITE_DATE		DateADTGetDatum(DATEVAL_NOEND)

/* Plan caches for inserting into history tables */
static HTAB *InsertHistoryPlanHash = NULL;

typedef struct InsertHistoryPlanEntry
{
	Oid			history_relid;	/* the hash key; must be first */
	char		schemaname[NAMEDATALEN];
	char		tablename[NAMEDATALEN];
	SPIPlanPtr	qplan;
} InsertHistoryPlanEntry;

static HTAB *
CreateInsertHistoryPlanHash(void)
{
	HASHCTL	ctl;

	ctl.keysize = sizeof(Oid);
	ctl.entrysize = sizeof(InsertHistoryPlanEntry);

	return hash_create("Insert History Hash", 16, &ctl, HASH_ELEM | HASH_BLOBS);
}

static void
GetPeriodColumnNames(Relation rel, char *period_name, char **start_name, char **end_name)
{
	int				ret;
	Datum			values[2];
	SPITupleTable  *tuptable;
	bool			is_null;
	Datum			dat;
	MemoryContext	mcxt = CurrentMemoryContext; /* The context outside of SPI */

	const char *sql =
		"SELECT p.start_column_name, p.end_column_name "
		"FROM periods.periods AS p "
		"WHERE (p.table_name, p.period_name) = ($1, $2)";
	static SPIPlanPtr qplan = NULL;

	if (SPI_connect() != SPI_OK_CONNECT)
		elog(ERROR, "SPI_connect failed");

	/*
	 * Query the periods table to get the start and end columns.
	 * Cache the plan if we haven't already.
	 */
	if (qplan == NULL)
	{
		Oid	types[2] = {OIDOID, NAMEOID};

		qplan = SPI_prepare(sql, 2, types);
		if (qplan == NULL)
			elog(ERROR, "SPI_prepare returned %s for %s",
				 SPI_result_code_string(SPI_result), sql);

		ret = SPI_keepplan(qplan);
		if (ret != 0)
			elog(ERROR, "SPI_keepplan returned %s", SPI_result_code_string(ret));
	}

	values[0] = ObjectIdGetDatum(rel->rd_id);
	values[1] = CStringGetDatum(period_name);
	ret = SPI_execute_plan(qplan, values, NULL, true, 0);
	if (ret != SPI_OK_SELECT)
		elog(ERROR, "SPI_execute returned %s", SPI_result_code_string(ret));

	/* Make sure we got one */
	if (SPI_processed == 0)
		ereport(ERROR,
				(errmsg("period \"%s\" not found on table \"%s\"",
						period_name,
						RelationGetRelationName(rel))));

	/* There is a unique constraint so there shouldn't be more than 1 row */
	Assert(SPI_processed == 1);

	/*
	 * Get the names from the result tuple.  We copy them into the original
	 * context so they don't get wiped out by SPI_finish().
	 */
	tuptable = SPI_tuptable;

	dat = SPI_getbinval(tuptable->vals[0], tuptable->tupdesc, 1, &is_null);
	*start_name = MemoryContextStrdup(mcxt, NameStr(*(DatumGetName(dat))));

	dat = SPI_getbinval(tuptable->vals[0], tuptable->tupdesc, 2, &is_null);
	*end_name = MemoryContextStrdup(mcxt, NameStr(*(DatumGetName(dat))));

	/* All done with SPI */
	if (SPI_finish() != SPI_OK_FINISH)
		elog(ERROR, "SPI_finish failed");
}

/*
 * Check if the only columns changed in an UPDATE are columns that the user is
 * excluding from SYSTEM VERSIONING. One possible use case for this is a
 * "last_login timestamptz" column on a user table.  Arguably, this column
 * should be in another table, but users have requested the feature so let's do
 * it.
 */
static bool
OnlyExcludedColumnsChanged(Relation rel, HeapTuple old_row, HeapTuple new_row)
{
	int				ret;
	int				i;
	Datum			values[1];
	TupleDesc		tupdesc = RelationGetDescr(rel);
	Bitmapset	   *excluded_attnums = NULL;
	MemoryContext	mcxt = CurrentMemoryContext; /* The context outside of SPI */

	const char *sql =
		"SELECT u.name "
		"FROM periods.system_time_periods AS stp "
		"CROSS JOIN unnest(stp.excluded_column_names) AS u (name) "
		"WHERE stp.table_name = $1";
	static SPIPlanPtr qplan = NULL;

	if (SPI_connect() != SPI_OK_CONNECT)
		elog(ERROR, "SPI_connect failed");

	/*
	 * Get the excluded column names.
	 * Cache the plan if we haven't already.
	 */
	if (qplan == NULL)
	{
		Oid	types[1] = {OIDOID};

		qplan = SPI_prepare(sql, 1, types);
		if (qplan == NULL)
			elog(ERROR, "SPI_prepare returned %s for %s",
				 SPI_result_code_string(SPI_result), sql);

		ret = SPI_keepplan(qplan);
		if (ret != 0)
			elog(ERROR, "SPI_keepplan returned %s", SPI_result_code_string(ret));
	}

	values[0] = ObjectIdGetDatum(rel->rd_id);
	ret = SPI_execute_plan(qplan, values, NULL, true, 0);
	if (ret != SPI_OK_SELECT)
		elog(ERROR, "SPI_execute returned %s", SPI_result_code_string(ret));

	/* Construct a bitmap of excluded attnums */
	if (SPI_processed > 0 && SPI_tuptable != NULL)
	{
		TupleDesc	spitupdesc = SPI_tuptable->tupdesc;
		bool		isnull;

		for (i = 0; i < SPI_processed; i++)
		{
			HeapTuple	tuple = SPI_tuptable->vals[i];
			Datum		attdatum;
			char	   *attname;
			int16		attnum;

			/* Get the attnum from the column name */
			attdatum = SPI_getbinval(tuple, spitupdesc, 1, &isnull);
			attname = NameStr(*(DatumGetName(attdatum)));
			attnum = SPI_fnumber(tupdesc, attname);

			/* Make sure it's valid (should always be) */
			if (attnum == SPI_ERROR_NOATTRIBUTE)
				ereport(ERROR,
						(errcode(ERRCODE_UNDEFINED_COLUMN),
						 errmsg("column \"%s\" does not exist", attname)));

			/* Just ignore system columns (should never happen) */
			if (attnum < 0)
				continue;

			/* Add it to the bitmap set */
			excluded_attnums = bms_add_member(excluded_attnums, attnum);
		}

		/*
		 * If we have excluded columns, move the bitmapset out of the SPI
		 * context.
		 */
		if (excluded_attnums != NULL)
		{
			MemoryContext spicontext = MemoryContextSwitchTo(mcxt);
			excluded_attnums = bms_copy(excluded_attnums);
			MemoryContextSwitchTo(spicontext);
		}
	}

	/* Don't need SPI anymore */
	if (SPI_finish() != SPI_OK_FINISH)
		elog(ERROR, "SPI_finish failed");

	/* If there are no excluded columns, then we're done */
	if (excluded_attnums == NULL)
		return false;

	for (i = 1; i <= tupdesc->natts; i++)
	{
		Datum	old_datum, new_datum;
		bool	old_isnull, new_isnull;
		int16	typlen;
		bool	typbyval;

		/* Ignore if deleted column */
		if (TupleDescAttr(tupdesc, i-1)->attisdropped)
			continue;

		/* Ignore if excluded column */
		if (bms_is_member(i, excluded_attnums))
			continue;

		old_datum = SPI_getbinval(old_row, tupdesc, i, &old_isnull);
		new_datum = SPI_getbinval(new_row, tupdesc, i, &new_isnull);

		/*
		 * If one value is NULL and other is not, then they are certainly not
		 * equal.
		 */
		if (old_isnull != new_isnull)
			return false;

		/* If both are NULL, they can be considered equal. */
		if (old_isnull)
			continue;

		/* Do a fairly strict binary comparison of the values */
		typlen = TupleDescAttr(tupdesc, i-1)->attlen;
		typbyval = TupleDescAttr(tupdesc, i-1)->attbyval;
		if (!datumIsEqual(old_datum, new_datum, typbyval, typlen))
			return false;
	}

	return true;
}

/*
 * Get the oid of the history table.  If this table does not have a system_time
 * period an error is raised.  If it doesn't have SYSTEM VERSIONING, then
 * InvalidOid is returned.
 */
static Oid
GetHistoryTable(Relation rel)
{
	int		ret;
	Datum	values[1];
	Oid		result;
	SPITupleTable  *tuptable;
	bool			is_null;

	const char *sql =
		"SELECT history_table_name::oid "
		"FROM periods.system_versioning AS sv "
		"WHERE sv.table_name = $1";
	static SPIPlanPtr qplan = NULL;

	if (SPI_connect() != SPI_OK_CONNECT)
		elog(ERROR, "SPI_connect failed");

	/*
	 * Check existence in system_versioning table.
	 * Cache the plan if we haven't already.
	 */
	if (qplan == NULL)
	{
		Oid	types[1] = {OIDOID};

		qplan = SPI_prepare(sql, 1, types);
		if (qplan == NULL)
			elog(ERROR, "SPI_prepare returned %s for %s",
				 SPI_result_code_string(SPI_result), sql);

		ret = SPI_keepplan(qplan);
		if (ret != 0)
			elog(ERROR, "SPI_keepplan returned %s", SPI_result_code_string(ret));
	}

	values[0] = ObjectIdGetDatum(rel->rd_id);
	ret = SPI_execute_plan(qplan, values, NULL, true, 0);
	if (ret != SPI_OK_SELECT)
		elog(ERROR, "SPI_execute returned %s", SPI_result_code_string(ret));

	/* Did we get one? */
	if (SPI_processed == 0)
	{
		if (SPI_finish() != SPI_OK_FINISH)
			elog(ERROR, "SPI_finish failed");
		return InvalidOid;
	}

	/* There is a unique constraint so there shouldn't be more than 1 row */
	Assert(SPI_processed == 1);

	/* Get oid from results */
	tuptable = SPI_tuptable;
	result = DatumGetObjectId(SPI_getbinval(tuptable->vals[0], tuptable->tupdesc, 1, &is_null));

	if (SPI_finish() != SPI_OK_FINISH)
		elog(ERROR, "SPI_finish failed");

	return result;
}

static Datum
GetRowStart(Oid typeid)
{
	switch (typeid)
	{
		case TIMESTAMPTZOID:
			return TRANSACTION_TSTZ;
		case TIMESTAMPOID:
			return TRANSACTION_TS;
		case DATEOID:
			return TRANSACTION_DATE;
		default:
			elog(ERROR, "unexpected type: %d", typeid);
			return 0;	/* keep compiler quiet */
	}
}

static Datum
GetRowEnd(Oid typeid)
{
	switch (typeid)
	{
		case TIMESTAMPTZOID:
			return INFINITE_TSTZ;
		case TIMESTAMPOID:
			return INFINITE_TS;
		case DATEOID:
			return INFINITE_DATE;
		default:
			elog(ERROR, "unexpected type: %d", typeid);
			return 0;	/* keep compiler quiet */
	}
}

static int
CompareWithCurrentDatum(Oid typeid, Datum value)
{
	switch (typeid)
	{
		case TIMESTAMPTZOID:
			return DatumGetInt32(DirectFunctionCall2(timestamp_cmp, value, TRANSACTION_TSTZ));

		case TIMESTAMPOID:
			return DatumGetInt32(DirectFunctionCall2(timestamp_cmp, value, TRANSACTION_TS));

		case DATEOID:
			return DatumGetInt32(DirectFunctionCall2(date_cmp, value, TRANSACTION_DATE));

		default:
			elog(ERROR, "unexpected type: %d", typeid);
			return 0;	/* keep compiler quiet */
	}
}

static int
CompareWithInfiniteDatum(Oid typeid, Datum value)
{
	switch (typeid)
	{
		case TIMESTAMPTZOID:
			return DatumGetInt32(DirectFunctionCall2(timestamp_cmp, value, INFINITE_TSTZ));

		case TIMESTAMPOID:
			return DatumGetInt32(DirectFunctionCall2(timestamp_cmp, value, INFINITE_TS));

		case DATEOID:
			return DatumGetInt32(DirectFunctionCall2(date_cmp, value, INFINITE_DATE));

		default:
			elog(ERROR, "unexpected type: %d", typeid);
			return 0;	/* keep compiler quiet */
	}
}

Datum
generated_always_as_row_start_end(PG_FUNCTION_ARGS)
{
	TriggerData	   *trigdata = castNode(TriggerData, fcinfo->context);
	const char	   *funcname = "generated_always_as_row_start_end";
	Relation		rel;
	HeapTuple		new_row;
	TupleDesc		new_tupdesc;
	Datum			values[2];
	bool			nulls[2];
	int				columns[2];
	char		   *start_name, *end_name;
	int16			start_num, end_num;
	Oid				typeid;

	/*
	 * Make sure this is being called as an BEFORE ROW trigger.  Note:
	 * translatable error strings are shared with ri_triggers.c, so resist the
	 * temptation to fold the function name into them.
	 */
	if (!CALLED_AS_TRIGGER(fcinfo))
		ereport(ERROR,
				(errcode(ERRCODE_E_R_I_E_TRIGGER_PROTOCOL_VIOLATED),
				 errmsg("function \"%s\" was not called by trigger manager",
						funcname)));

	if (!TRIGGER_FIRED_BEFORE(trigdata->tg_event) ||
		!TRIGGER_FIRED_FOR_ROW(trigdata->tg_event))
		ereport(ERROR,
				(errcode(ERRCODE_E_R_I_E_TRIGGER_PROTOCOL_VIOLATED),
				 errmsg("function \"%s\" must be fired BEFORE ROW",
						funcname)));

	/* Get Relation information */
	rel = trigdata->tg_relation;
	new_tupdesc = RelationGetDescr(rel);

	/* Get the new data that was inserted/updated */
	if (TRIGGER_FIRED_BY_INSERT(trigdata->tg_event))
		new_row = trigdata->tg_trigtuple;
	else if (TRIGGER_FIRED_BY_UPDATE(trigdata->tg_event))
	{
		HeapTuple old_row;

		old_row = trigdata->tg_trigtuple;
		new_row = trigdata->tg_newtuple;

		/* Don't change anything if only excluded columns are being updated. */
		if (OnlyExcludedColumnsChanged(rel, old_row, new_row))
			return PointerGetDatum(new_row);
	}
	else
	{
		ereport(ERROR,
				(errcode(ERRCODE_E_R_I_E_TRIGGER_PROTOCOL_VIOLATED),
				 errmsg("function \"%s\" must be fired for INSERT or UPDATE",
						funcname)));
		new_row = NULL;			/* keep compiler quiet */
	}

	GetPeriodColumnNames(rel, "system_time", &start_name, &end_name);

	/* Get the column numbers and type */
	start_num = SPI_fnumber(new_tupdesc, start_name);
	end_num = SPI_fnumber(new_tupdesc, end_name);
	typeid = SPI_gettypeid(new_tupdesc, start_num);

	columns[0] = start_num;
	values[0] = GetRowStart(typeid);
	nulls[0] = false;
	columns[1] = end_num;
	values[1] = GetRowEnd(typeid);
	nulls[1] = false;
#if (PG_VERSION_NUM < 100000)
	new_row = SPI_modifytuple(rel, new_row, 2, columns, values, nulls);
#else
	new_row = heap_modify_tuple_by_cols(new_row, new_tupdesc, 2, columns, values, nulls);
#endif

	return PointerGetDatum(new_row);
}

static void
insert_into_history(Relation history_rel, HeapTuple history_tuple)
{
	InsertHistoryPlanEntry   *hentry;
	bool		found;
	char	   *schemaname = SPI_getnspname(history_rel);
	char	   *tablename = SPI_getrelname(history_rel);
	Oid			history_relid = history_rel->rd_id;
	Datum		value;
	int			ret;

	if (SPI_connect() != SPI_OK_CONNECT)
		elog(ERROR, "SPI_connect failed");

	if (!InsertHistoryPlanHash)
		InsertHistoryPlanHash = CreateInsertHistoryPlanHash();

	/* Fetch the cached plan */
	hentry = (InsertHistoryPlanEntry *) hash_search(
			InsertHistoryPlanHash,
			&history_relid,
			HASH_ENTER,
			&found);

	/* If we didn't find it or the name changed, re-plan it */
	if (!found ||
		!strcmp(hentry->schemaname, schemaname) ||
		!strcmp(hentry->tablename, tablename))
	{
		StringInfo	buf = makeStringInfo();
		Oid			type = HeapTupleHeaderGetTypeId(history_tuple->t_data);

		appendStringInfo(buf, "INSERT INTO %s VALUES (($1).*)",
				quote_qualified_identifier(schemaname, tablename));

		hentry->history_relid = history_relid;
		strlcpy(hentry->schemaname, schemaname, sizeof(hentry->schemaname));
		strlcpy(hentry->tablename, tablename, sizeof(hentry->tablename));
		hentry->qplan = SPI_prepare(buf->data, 1, &type);
		if (hentry->qplan == NULL)
			elog(ERROR, "SPI_prepare returned %s for %s",
				 SPI_result_code_string(SPI_result), buf->data);

		ret = SPI_keepplan(hentry->qplan);
		if (ret != 0)
			elog(ERROR, "SPI_keepplan returned %s", SPI_result_code_string(ret));
	}

	/* Do the INSERT */
	value = HeapTupleGetDatum(history_tuple);
	ret = SPI_execute_plan(hentry->qplan, &value, NULL, false, 0);
	if (ret != SPI_OK_INSERT)
		elog(ERROR, "SPI_execute returned %s", SPI_result_code_string(ret));

	if (SPI_finish() != SPI_OK_FINISH)
		elog(ERROR, "SPI_finish failed");
}

Datum
write_history(PG_FUNCTION_ARGS)
{
	TriggerData	   *trigdata = castNode(TriggerData, fcinfo->context);
	const char	   *funcname = "write_history";
	Relation		rel;
	HeapTuple		old_row, new_row;
	TupleDesc		tupledesc;
	char		   *start_name, *end_name;
	int16			start_num, end_num;
	Oid				typeid;
	bool			is_null;
	Oid				history_id;
	int				cmp;
	bool			only_excluded_changed = false;

	/*
	 * Make sure this is being called as an AFTER ROW trigger.  Note:
	 * translatable error strings are shared with ri_triggers.c, so resist the
	 * temptation to fold the function name into them.
	 */
	if (!CALLED_AS_TRIGGER(fcinfo))
		ereport(ERROR,
				(errcode(ERRCODE_E_R_I_E_TRIGGER_PROTOCOL_VIOLATED),
				 errmsg("function \"%s\" was not called by trigger manager",
						funcname)));

	if (!TRIGGER_FIRED_AFTER(trigdata->tg_event) ||
		!TRIGGER_FIRED_FOR_ROW(trigdata->tg_event))
		ereport(ERROR,
				(errcode(ERRCODE_E_R_I_E_TRIGGER_PROTOCOL_VIOLATED),
				 errmsg("function \"%s\" must be fired AFTER ROW",
						funcname)));

	/* Get Relation information */
	rel = trigdata->tg_relation;
	tupledesc = RelationGetDescr(rel);

	/* Get the old data that was updated/deleted */
	if (TRIGGER_FIRED_BY_INSERT(trigdata->tg_event))
	{
		old_row = NULL;			/* keep compiler quiet */
		new_row = trigdata->tg_trigtuple;
	}
	else if (TRIGGER_FIRED_BY_UPDATE(trigdata->tg_event))
	{
		old_row = trigdata->tg_trigtuple;
		new_row = trigdata->tg_newtuple;

		/* Did only excluded columns change? */
		only_excluded_changed = OnlyExcludedColumnsChanged(rel, old_row, new_row);
	}
	else if (TRIGGER_FIRED_BY_DELETE(trigdata->tg_event))
	{
		old_row = trigdata->tg_trigtuple;
		new_row = NULL;			/* keep compiler quiet */
	}
	else
	{
		ereport(ERROR,
				(errcode(ERRCODE_E_R_I_E_TRIGGER_PROTOCOL_VIOLATED),
				 errmsg("function \"%s\" must be fired for INSERT or UPDATE or DELETE",
						funcname)));
		old_row = NULL;			/* keep compiler quiet */
		new_row = NULL;			/* keep compiler quiet */
	}

	GetPeriodColumnNames(rel, "system_time", &start_name, &end_name);

	/* Get the column numbers and type */
	start_num = SPI_fnumber(tupledesc, start_name);
	end_num = SPI_fnumber(tupledesc, end_name);
	typeid = SPI_gettypeid(tupledesc, start_num);

	/*
	 * Validate that the period columns haven't been modified.  This can happen
	 * with a trigger executed after generated_always_as_row_start_end().
	 */
	if (TRIGGER_FIRED_BY_INSERT(trigdata->tg_event) ||
		(TRIGGER_FIRED_BY_UPDATE(trigdata->tg_event) && !only_excluded_changed))
	{
		Datum	start_datum = SPI_getbinval(new_row, tupledesc, start_num, &is_null);
		Datum	end_datum = SPI_getbinval(new_row, tupledesc, end_num, &is_null);

		if (CompareWithCurrentDatum(typeid, start_datum) != 0)
			ereport(ERROR,
					(errcode(ERRCODE_GENERATED_ALWAYS),
					 errmsg("cannot insert or update column \"%s\"", start_name),
					 errdetail("Column \"%s\" is GENERATED ALWAYS AS ROW START", start_name)));

		if (CompareWithInfiniteDatum(typeid, end_datum) != 0)
			ereport(ERROR,
					(errcode(ERRCODE_GENERATED_ALWAYS),
					 errmsg("cannot insert or update column \"%s\"", end_name),
					 errdetail("Column \"%s\" is GENERATED ALWAYS AS ROW END", end_name)));

		/*
		 * If this is an INSERT, then we're done because there is no history to
		 * write.
		 */
		if (TRIGGER_FIRED_BY_INSERT(trigdata->tg_event))
			return PointerGetDatum(NULL);
	}

	/* If only excluded columns have changed, don't write history. */
	if (only_excluded_changed)
		return PointerGetDatum(NULL);

	/* Compare the OLD row's start with the transaction start */
	cmp = CompareWithCurrentDatum(typeid,
			SPI_getbinval(old_row, tupledesc, start_num, &is_null));

	/*
	 * Don't do anything more if the start time is still the same.
	 *
	 * DELETE: SQL:2016 13.4 GR 15)a)iii)2)
	 * UPDATE: SQL:2016 15.13 GR 9)a)iii)2)
	 */
	if (cmp == 0)
		return PointerGetDatum(NULL);

	/*
	 * There is a weird case in READ UNCOMMITTED and READ COMMITTED where a
	 * transaction can UPDATE/DELETE a row created by a transaction that
	 * started later.  In effect, system-versioned tables must be run at the
	 * SERIALIZABLE level and so if we come across such an anomaly, we give an
	 * invalid row version error, per spec.
	 *
	 * DELETE: SQL:2016 13.4 GR 15)a)iii)1)
	 * UPDATE: SQL:2016 15.13 GR 9)a)iii)1)
	 */
	if (cmp > 0)
		ereport(ERROR,
				(errcode(ERRCODE_INVALID_ROW_VERSION),
				 errmsg("invalid row version"),
				 errdetail("The row being updated or deleted was created after this transaction started."),
				 errhint("The transaction might succeed if retried.")));

	/*
	 * If this table does not have SYSTEM VERSIONING, there is nothing else to
	 * be done.
	 */
	history_id = GetHistoryTable(rel);
	if (OidIsValid(history_id))
	{
		Relation	history_rel;
		TupleDesc	history_tupledesc;
		HeapTuple	history_tuple;
		int16		history_end_num;
		TupleConversionMap   *map;
		Datum	   *values;
		bool	   *nulls;

		/* Open the history table for inserting */
		history_rel = table_open(history_id, RowExclusiveLock);
		history_tupledesc = RelationGetDescr(history_rel);
		history_end_num = SPI_fnumber(history_tupledesc, end_name);

		/*
		 * We may have to convert the tuple structure between the table and the
		 * history table.
		 *
		 * See https://github.com/xocolatl/periods/issues/5
		 */
#if (PG_VERSION_NUM < 130000)
		map = convert_tuples_by_name(tupledesc, history_tupledesc, gettext_noop("could not convert row type"));
#else
		map = convert_tuples_by_name(tupledesc, history_tupledesc);
#endif
		if (map != NULL)
		{
#if (PG_VERSION_NUM < 120000)
			history_tuple = do_convert_tuple(old_row, map);
#else
			history_tuple = execute_attr_map_tuple(old_row, map);
#endif
			free_conversion_map(map);
		}
		else
		{
			history_tuple = old_row;

			/*
			 * Use the main table's tupledesc if there is no map so that
			 * missing attributes are filled in.  This corrects for bug #16242
			 * which was found by this very problem.
			 */
			history_tupledesc = tupledesc;
		}

		/* Build the new tuple for the history table */
		values = (Datum *) palloc(history_tupledesc->natts * sizeof(Datum));
		nulls = (bool *) palloc(history_tupledesc->natts * sizeof(bool));

		/* Modify the historical ROW END on the fly */
		heap_deform_tuple(history_tuple, history_tupledesc, values, nulls);
		values[history_end_num-1] = GetRowStart(typeid);
		nulls[history_end_num-1] = false;
		history_tuple = heap_form_tuple(history_tupledesc, values, nulls);

		pfree(values);
		pfree(nulls);

		/* INSERT the row */
		insert_into_history(history_rel, history_tuple);

		/* Keep the lock until end of transaction */
		table_close(history_rel, NoLock);
	}

	return PointerGetDatum(NULL);
}