Estoy tratando de entender los planes físicos en spark pero no entiendo algunas partes porque parecen diferentes de los rdbms tradicionales. Por ejemplo, en este plan a continuación, es un plan sobre una consulta sobre una tabla de colmena. La consulta es esta:

select
        l_returnflag,
        l_linestatus,
        sum(l_quantity) as sum_qty,
        sum(l_extendedprice) as sum_base_price,
        sum(l_extendedprice * (1 - l_discount)) as sum_disc_price,
        sum(l_extendedprice * (1 - l_discount) * (1 + l_tax)) as sum_charge,
        avg(l_quantity) as avg_qty,
        avg(l_extendedprice) as avg_price,
        avg(l_discount) as avg_disc,
        count(*) as count_order
    from
        lineitem
    where
        l_shipdate <= '1998-09-16'
    group by
        l_returnflag,
        l_linestatus
    order by
        l_returnflag,
        l_linestatus;


== Physical Plan ==
Sort [l_returnflag#35 ASC,l_linestatus#36 ASC], true, 0
+- ConvertToUnsafe
   +- Exchange rangepartitioning(l_returnflag#35 ASC,l_linestatus#36 ASC,200), None
      +- ConvertToSafe
         +- TungstenAggregate(key=[l_returnflag#35,l_linestatus#36], functions=[(sum(l_quantity#31),mode=Final,isDistinct=false),(sum(l_extendedpr#32),mode=Final,isDistinct=false),(sum((l_extendedprice#32 * (1.0 - l_discount#33))),mode=Final,isDistinct=false),(sum(((l_extendedprice#32 * (1.0l_discount#33)) * (1.0 + l_tax#34))),mode=Final,isDistinct=false),(avg(l_quantity#31),mode=Final,isDistinct=false),(avg(l_extendedprice#32),mode=Fl,isDistinct=false),(avg(l_discount#33),mode=Final,isDistinct=false),(count(1),mode=Final,isDistinct=false)], output=[l_returnflag#35,l_linestatus,sum_qty#0,sum_base_price#1,sum_disc_price#2,sum_charge#3,avg_qty#4,avg_price#5,avg_disc#6,count_order#7L])
            +- TungstenExchange hashpartitioning(l_returnflag#35,l_linestatus#36,200), None
               +- TungstenAggregate(key=[l_returnflag#35,l_linestatus#36], functions=[(sum(l_quantity#31),mode=Partial,isDistinct=false),(sum(l_exdedprice#32),mode=Partial,isDistinct=false),(sum((l_extendedprice#32 * (1.0 - l_discount#33))),mode=Partial,isDistinct=false),(sum(((l_extendedpri32 * (1.0 - l_discount#33)) * (1.0 + l_tax#34))),mode=Partial,isDistinct=false),(avg(l_quantity#31),mode=Partial,isDistinct=false),(avg(l_extendedce#32),mode=Partial,isDistinct=false),(avg(l_discount#33),mode=Partial,isDistinct=false),(count(1),mode=Partial,isDistinct=false)], output=[l_retulag#35,l_linestatus#36,sum#64,sum#65,sum#66,sum#67,sum#68,count#69L,sum#70,count#71L,sum#72,count#73L,count#74L])
                  +- Project [l_discount#33,l_linestatus#36,l_tax#34,l_quantity#31,l_extendedprice#32,l_returnflag#35]
                     +- Filter (l_shipdate#37 <= 1998-09-16)
                        +- HiveTableScan [l_discount#33,l_linestatus#36,l_tax#34,l_quantity#31,l_extendedprice#32,l_shipdate#37,l_returnflag#35], astoreRelation default, lineitem, None

Por lo que entiendo en el plan es:

Primero comienza con un escaneo de tabla Hive

Luego se filtra usando donde la condición

Luego proyecte para obtener las columnas que queremos

Entonces TungstenAggregate?

Entonces TungstenExchange?

Entonces TungstenAggregate nuevamente?

Entonces ConvertToSafe?

Luego ordena el resultado final

Pero no entiendo los 4, 5, 6 y 7 pasos. ¿Sabes lo que son? Estoy buscando información sobre esto para poder entender el plan, pero no encuentro nada concreto.