tdfs4ds.serve.bootstrap¶
Runtime bootstrap for the tdfs4ds API server.
Establishes the Teradata connection, lets tdfs4ds load its own
TDFS4DS_* configuration (SCHEMA, DATA_DOMAIN, INSTRUCT_MODEL_,
EMBEDDING_MODEL_, CHROMA_*, VECTOR_STORE_BACKEND …), and optionally
pre-builds the consumer-agent vector index so the first SEARCH does not pay a
cold build.
The Teradata connection is not covered by tdfs4ds/config.py — that
module loads feature-store / LLM / vector-store settings but never opens a
database context. This bootstrap owns the connection and reads it directly
from the environment:
TDFS4DS_TD_HOST Teradata host (required)
TDFS4DS_TD_USERNAME Teradata user (required for a password logon)
TDFS4DS_TD_PASSWORD Teradata password (required for a password logon)
TDFS4DS_TD_LOGDATA Token logon instead of a password, e.g. 'token=<jwt>'
(used only when no password is set)
TDFS4DS_TD_LOGMECH Logon mechanism (optional, e.g. LDAP / TD2; JWT for a token)
TDFS4DS_TD_DATABASE Default database for the session (optional)
TDFS4DS_TD_ENCRYPT Enable wire encryption — 'true' or 'false' (optional)
Server-specific flags:
TDFS4DS_SERVE_BUILD_INDEX 'true' (default) builds the vector index at
startup; set 'false' for a QO-only deployment.
module-attribute
¶
_FORM_KEYS = {'host': 'TDFS4DS_TD_HOST', 'username': 'TDFS4DS_TD_USERNAME', 'password': 'TDFS4DS_TD_PASSWORD', 'logmech': 'TDFS4DS_TD_LOGMECH', 'database': 'TDFS4DS_TD_DATABASE', 'encryptdata': 'TDFS4DS_TD_ENCRYPT'}
Small status payload for /health and the UI status dot.
create_context kwargs for a token logon, or None for the password path.
Only when no password is available and TDFS4DS_TD_LOGDATA is set: the
shape a platform that authenticates with a token gives (AI Studio:
logmech='JWT', logdata='token=…'). The kwargs come from
:func:tdfs4ds.utils.reconnect._connect_kwargs, the one place that maps
TDFS4DS_TD_* onto a connection, so the startup logon and the automatic
reconnect agree. With a password set this returns None and the historical
path runs unchanged.
Open a teradataml context from TDFS4DS_TD_* environment variables.
Returns the teradataml context object. Raises RuntimeError with a
clear message when the mandatory connection variables are missing.
A failed logon, with the errors it was raised from.
teradataml raises TDML_2000 Failed to connect to Teradata Vantage from
the driver's error, which is the one naming the fault (bad password, LDAP
refusal, unknown host, encryption setting). str(exc) keeps only the
first, so the Connect page said nothing useful. Every message of the chain
(__cause__, else __context__), whole, without the driver's Go frames.
connect_teradata(host=None, username=None, password=None, logmech=None, database=None, encrypt=None, build_index=None)
Open (or re-open) the Teradata context and run post-connection setup.
Reusable by both the startup bootstrap and the in-app Connect page. Any
credential left None falls back to the matching TDFS4DS_TD_* env
var. On success, sets the module CONNECTED flag, applies tdfs4ds config
and (optionally) builds the vector index. On failure, CONNECTED stays
False and LAST_CONNECT_ERROR records the reason — the caller decides
whether to raise.
Returns {"connected": bool, "error": str|None}.
Apply tdfs4ds config (already loaded on import) and build the index.
Use the session's default database as the feature store when it holds one.
import tdfs4ds takes SCHEMA from the teradataml context, but the app
imports tdfs4ds before it connects, so a logon with database= (the Connect
page's Default database, or TDFS4DS_TD_DATABASE) left SCHEMA empty and
every panel failed on FROM .FS_FEATURE_CATALOG (seen at a site). Adopted only
when that database holds a complete feature store; returns its name, else ''.
Never fails the startup.
Attach the read-only stores named in the configuration properly.
TDFS4DS_SCHEMA_READONLY (or JSON schema_readonly) only sets
tdfs4ds.SCHEMA_READONLY to plain names; nothing checked them or probed
what they offer until a store switch in Parameters. extend_read_stores
does both. Never fails the startup.
Status-bar key of a store's index build (one build per store at a time).
Build the vector index, one job per store, shown in the status bar.
stores: database names or descriptors (default: every attached store, the primary first). Returns the job keys.
In the background only when it is safe: on Chroma, with the
TDFS4DS_TD_* credentials to open a connection of its own
(vector_index.own_connection), since a background read on teradataml's
single shared connection breaks it for the whole process (seen live). The
Teradata Vector Store backend works through teradataml throughout, so it
builds in the calling thread, as before 0.3.3.15, for the primary only; so
does Chroma when no connection of its own can be opened.
'objects 3 embedded / 40 unchanged, …' from a Chroma build summary ('' otherwise).
Idempotent, NON-FATAL server bootstrap.
Attempts the Teradata connection from TDFS4DS_TD_* env vars. If it fails
(missing creds or unreachable host) the app still boots disconnected —
the shell, /health and the in-app Connect page load, and the user can
connect at runtime via POST /ui/connect. Safe to call more than once.