TiDB Lightning Glossary
This page explains the special terms used in TiDB Lightning's logs, monitoring, configurations, and documentation.
A
Analyze
An operation to rebuild the statistics information of a TiDB table, which is running the ANALYZE TABLE
statement.
Because TiDB Lightning imports data without going through TiDB, the statistics information is not automatically updated. Therefore, TiDB Lightning explicitly analyzes every table after importing. This step can be omitted by setting the post-restore.analyze
configuration to false
.
AUTO_INCREMENT_ID
Every table has an associated AUTO_INCREMENT_ID
counter to provide the default value of an auto-incrementing column. In TiDB, this counter is additionally used to assign row IDs.
Because TiDB Lightning imports data without going through TiDB, the AUTO_INCREMENT_ID
counter is not automatically updated. Therefore, TiDB Lightning explicitly alters AUTO_INCREMENT_ID
to a valid value. This step is always performed, even if the table has no AUTO_INCREMENT
columns.
B
Black-white list
A configuration list that specifies which tables to be imported and which should be excluded.
See TiDB Lightning Table Filter for details.
C
Checkpoint
TiDB Lightning continuously saves its progress into a local file or a remote database while importing. This allows it to resume from an intermediate state should it crashes in the process. See the Checkpoints section for details.
Checksum
In TiDB Lightning, the checksum of a table is a set of 3 numbers calculated from the content of each KV pair in that table. These numbers are respectively:
- the number of KV pairs,
- total length of all KV pairs, and
- the bitwise-XOR of CRC-64-ECMA value each pair.
TiDB Lightning validates the imported data by comparing the local and remote checksums of every table. The program would stop if any pair does not match. You can skip this check by setting the post-restore.checksum
configuration to false
.
See also the Troubleshooting guide for how to properly handle checksum mismatch.
Chunk
Equivalent to a single file in the data source.
Compaction
An operation that merges multiple small SST files into one large SST file, and cleans up the deleted entries. TiKV automatically compacts data in background while TiDB Lightning is importing.
See RocksDB's wiki page on Compaction for its technical details.
D
Data engine
An engine for sorting actual row data.
When a table is very large, its data is placed into multiple data engines to improve task pipelining and save space of TiKV Importer. By default, a new data engine is opened for every 100 GB of SQL data, which can be configured through the mydumper.batch-size
setting.
TiDB Lightning processes multiple data engines concurrently. This is controlled by the lightning.table-concurrency
setting.
E
Engine
In TiKV Importer, an engine is a RocksDB instance for sorting KV pairs.
TiDB Lightning transfers data to TiKV Importer through engines. It first opens an engine, sends KV pairs to it (with no particular order), and finally closes the engine. The engine sorts the received KV pairs after it is closed. These closed engines can then be further uploaded to the TiKV stores for ingestion.
Engines use TiKV Importer's import-dir
as temporary storage, which are sometimes referred to as "engine files".
See also data engine and index engine.
I
Import mode
A configuration that optimizes TiKV for writing at the cost of degraded read speed and space usage.
TiDB Lightning automatically switches to and off the import mode while running. However, if TiKV gets stuck in import mode, you can use tidb-lightning-ctl
to force revert to normal mode.
Index engine
An engine for sorting indices.
Regardless of number of indices, every table is associated with exactly one index engine.
TiDB Lightning processes multiple index engines concurrently. This is controlled by the lightning.index-concurrency
setting. Since every table has exactly one index engine, this also configures the maximum number of tables to process at the same time.
Ingest
An operation which inserts the entire content of an SST file into the RocksDB (TiKV) store.
Ingestion is a very fast operation compared with inserting KV pairs one by one. This operation is the determinant factor for the performance of TiDB Lightning.
See RocksDB's wiki page on Creating and Ingesting SST files for its technical details.
K
KV pair
Abbreviation of "key-value pair".
KV encoder
A routine which parses SQL or CSV rows to KV pairs. Multiple KV encoders run in parallel to speed up processing.
L
Local checksum
The checksum of a table calculated by TiDB Lightning itself before sending the KV pairs to TiKV Importer.
N
Normal mode
The mode where import mode is disabled.
P
Post-processing
The period of time after the entire data source is parsed and sent to TiKV Importer. TiDB Lightning is waiting for TiKV Importer to upload and ingest the SST files.
R
Remote checksum
The checksum of a table calculated by TiDB after it has been imported.
S
Scattering
An operation that randomly reassigns the leader and the peers of a Region. Scattering ensures that the imported data are distributed evenly among TiKV stores. This reduces stress on PD.
Splitting
An engine is typically very large (around 100 GB), which is not friendly to TiKV if treated as a single region. TiKV Importer splits an engine into multiple small SST files (configurable by TiKV Importer's import.region-split-size
setting) before uploading.
SST file
SST is the abbreviation of "sorted string table". An SST file is RocksDB's (and thus TiKV's) native storage format of a collection of KV pairs.
TiKV Importer produces SST files from a closed engine. These SST files are uploaded and then ingested into TiKV stores.