Teradata SQL, where the DDL is the physical design.

In every other dialect on our list, CREATE TABLE declares columns. In Teradata it also declares duplicate-row semantics, the fallback copy policy, whether the table is journaled before and after, the data block size in bytes, the checksum level, the merge block ratio, the hashing map, the primary index and the compression values for individual columns. Then it declares the columns.

CREATE MULTISET TABLE prodeiw.SB_TRANS_EVENT ,FALLBACK ,
     NO BEFORE JOURNAL, NO AFTER JOURNAL,
     DATABLOCKSIZE = 524288 BYTES, CHECKSUM = DEFAULT,
     DEFAULT MERGEBLOCKRATIO, MAP = TD_MAP1
     (
      SB_Event_Id INTEGER NOT NULL,
      SB_Event_Processing_Dt DATE FORMAT 'YYYY-MM-DD',
      SB_Activity_Cd INTEGER COMPRESS (10251 ,10000 ,-10 ),
      SB_Host_Val CHAR(10) CHARACTER SET LATIN
          NOT CASESPECIFIC
     )
PRIMARY INDEX ( SB_Event_Id );

teradata_sql_compilation_unit
└── teradata_sql_statement
    ├── teradata_sql_table_multiplicity MULTISET
    ├── teradata_sql_fallback_table_option FALLBACK
    ├── teradata_sql_before_journal_table_option
    ├── teradata_sql_data_block_size_table_option
    │     524288 BYTES
    ├── teradata_sql_checksum_table_option DEFAULT
    ├── teradata_sql_merge_block_ratio_table_option
    ├── teradata_sql_map_table_option TD_MAP1
    ├── ⋯ column definitions
    └── teradata_sql_primary_index SB_Event_Id

BTEQ: what actually happens to the dot-commands

This is the first question every Teradata prospect asks, and it deserves an exact answer rather than a reassuring one.

BTEQ_COMMAND
    : {this.getCharPositionInLine() == 0}?
      (PERIOD|EQ) .*? SEMICOLON? '\n' -> channel(HIDDEN)
    ;

FastLoad, MultiLoad, TPump and TPT are a different matter: they are separate utility languages with their own control syntax, and they are out of scope — no token, no rule, no fixture. If your corpus is mostly MultiLoad scripts, this is not the product, and we would rather tell you on the first call.

Four constructs that decide whether a parser survives a real warehouse export. None of them is exotic Teradata; all four are in the first thousand lines of any DDL dump.

  cd INTEGER COMPRESS (10251 ,10000 ,-10 ),
  val CHAR(10) COMPRESS
      ('0001      ','SSVS01    ')

SELECT TOP 10 acct, amt,
       RANK() OVER (ORDER BY amt DESC) r
  FROM ledger
QUALIFY r <= 3;

LOCKING ROW FOR ACCESS
SELECT acct, bal FROM prodeiw.balances;

REPLACE VIEW prodeiw.v_open AS
SELECT NORMALIZE ON MEETS OR OVERLAPS
       acct, valid_period FROM prodeiw.spans;

Coverage, by construct

We do not publish a certified Teradata release range, and we are not going to invent one. What can be said is what the grammar demonstrably accepts: the MAP = TD_MAP1 clause in the shipped corpus is a 16.10-and-later feature, so the grammar is at least that modern. Every release of the engine ships its own example set, and that set is the honest record of what each version added.

We license five SQL-family engines. This is what the first page of a Teradata DDL export looks like, and exactly one of the five reads it.

LOCKING ROW FOR ACCESS
CREATE MULTISET TABLE prodeiw.SB_TRANS_EVENT ,FALLBACK ,
     NO BEFORE JOURNAL, NO AFTER JOURNAL,
     DATABLOCKSIZE = 524288 BYTES, CHECKSUM = DEFAULT,
     DEFAULT MERGEBLOCKRATIO, MAP = TD_MAP1
     (
      SB_Event_Id INTEGER NOT NULL,
      SB_File_Number SMALLINT NOT NULL,
      SB_Event_Processing_Dt DATE FORMAT 'YYYY-MM-DD' NOT NULL,
      SB_Activity_Type_Cd INTEGER NOT NULL COMPRESS (10251 ,10000 ,-10 ),
      SB_Host_Activity_Val CHAR(10) CHARACTER SET LATIN
          NOT CASESPECIFIC COMPRESS ('0001      ','SSVS01    ')
     )
PRIMARY INDEX ( SB_File_Number );

COLLECT STATISTICS USING SAMPLE 10 PERCENT
   COLUMN (SB_Event_Processing_Dt) ON prodeiw.SB_TRANS_EVENT;

You do not have to take this on trust. Paste a statement of your own into the Strumenta Playground, pick Teradata SQL, and read the tree the engine actually produces — including the issue list when it does not like something.

A jar, a Python package on top of it, the generated documents that describe the tree, and a person to write to. It runs inside your own network, with no connection back to us and none to the warehouse the DDL came from.

For Oracle or Transact-SQL there is a free public grammar to start from, and it is a reasonable starting point. For Teradata there is not: the platform is a closed commercial product and its documentation sits behind a login. The realistic free option is a generic ANSI SQL grammar, so it is worth being concrete about where that stops.

Writing the physical-design DDL alone is weeks of careful reading against documentation you have to have access to, and getting the BTEQ column-0 rule right is the kind of detail that only surfaces when a real customer script fails. The method behind this work has a name: read the Chisel Method, or Parsers and transpilers for the three ways into it.

Six things we would rather tell you now than have you find in week three. Two of them lose deals, and they are still here.