Difference between revisions of "AMI Historical Database"

From 3forge Documentation
Jump to navigation Jump to search
Line 601: Line 601:
  
 
[[File:FileDiretory2.png|750px]]<br>
 
[[File:FileDiretory2.png|750px]]<br>
 +
 +
If we go to '''''/ami/amione/hdb''''' directory, we will see all the folders, named as
 +
'''''Your_Historical_Table_Name.3fhtab''''' , where all the rows from the table are stored.<br>
 +
[[File:FolderDir.png|750px]]

Revision as of 11:29, 30 May 2023

Introduction to AMI Historical Database

The AMI Historical Table is a columnar-based on-disk table engine with support for partitioning. The introduction section will briefly describe the core concepts of the AMI Historical database: Partitioning, Data storage and optimizations.

Partitioning

Partitioning is used to organize data for performance. To understand why partitioning is critical for large datasets, let’s first consider a non-partitioned standard table: as the number or rows increases, the performance (time to insert, update, query and alter) will decline. Partitioning provides an elegant solution by breaking up rows into separate isolated “virtual tables” called partitions. Once data is broken into partitions, the row count of one partition will not affect the performance of other partitions. In order to dictate how rows are organized into partitions, the AMI Historical Table introduces a special PARTITION column type. The value in the PARTITION column(s) determines which partition a given row will be assigned to. In other words, all rows for a given partition will have the same value in the PARTITION column(s).

By way of example, data is often partitioned by a date: Imagine we have the following historical table, partitioned by the column “date”
Partition dynamic0.png
When inserting row(s) whose date match(es) the existing partition key(s), the query engine will automatically navigate to the corresponding partition according to the partition key and append the rows accordingly.
Partition dynamic1.png
Note that adding another day of data will automatically create a new partition without affecting existing days’ data: For example, if we want to insert two new rows into the table whose “Date” == 20221203
Partition dynamic2.png
To create a historical table add USE PersistEngine="HISTORICAL" and include the PARTITION keyword after the column(s) that the rows should be partitioned by. Ex:

CREATE PUBLIC TABLE PartitionDemo(date LONG PARTITION, symbol STRING, qty INTEGER, comments STRING) use PersistEngine="HISTORICAL";

The file structure of AMIHDB is such that each table has its own directory and each partition is assigned its own subdirectory. Each column and index has its own file(s) per partition directory.

<amiroot>/hdb/Table_Name.3fhtab/Partition_Id.dfhpar/<Files within a partition>

● Use ami.hdb.root.dir property to override location of hdb root
● Use ami.hdb.filehandles.max property to control how many files can be open at a time
● Table_Name - is the name of the table
● Partition_Id - An auto-incrementing id (starting at 0001)
● Files within a partition:
○ KEY.3fhkey - This stores the (plaintext) key for this partition
○ ColumnName.3fhcol - Stores data for the column
○ IndexName.3fhidx - Stores index data

This paradigm is designed for fast bulk INSERTs, fast random SELECTs and fast DELETES of entire partitions. UPDATES, and sparse DELETES are possible but will operate slower

Optimized for Schema Alterations

Each partition is a self-contained set of data. This means that altering a schema does not impact the way existing partitions are saved on disk. Instead, AMIHDB will map the “old” partitions using the “old” schema to the current schema definition at query time. This applies for ADD, DROP and MODIFY of columns.

Important Note: The PARTITION columns are the only columns that can not be altered after the table is created.

Data storage types and optimizations

The AMI historical database has different data storage types designed for data of various natures and characteristics, such as length, cardinality and partition keys. While continuously inserting, the nature of the data in the table is unpredictable so a more flexible but less efficient strategy is used for storage. Therefore, after the partition is completed, it is important to optimize the storage strategy to adapt to the existing data to improve query performance and minimize disk storage. Data storage types and optimizations will be further discussed in Section 2: Data types and storage types, and Section 4.1:Optimization on existing data.

PARTITION MODES

As a historical database, it is important to have a mode for fast insert-on-demand data. In AMIHDB, when a new partition is created it will have a status of IN_APPEND_MODE/NOT_OPTIMIZED. Then, after that table is optimized, the partition will become optimized for data compression and query times.

There are two ways to force optimize partitions:
(a). By calling CALL __OPTIMIZE_HISTORICIAL_TABLE(“Your_Historical_Table_Name”)
(b). By restarting AMI Center

You can use the DIAGNOSE TABLE <Your_Hitorical_Table_Name> command to view the optimized status of each partition:
Before commit:
Mode1.png
After commit:
Mode2.png

Data Types and Storage Type

AMIHDB supports the same data types as the AMI Real-time database:
BYTE, SHORT, INT, LONG, FLOAT, DOUBLE, CHAR, UTC, UTCN, UUID, COMPLEX, BINARY, STRING, BIGINT, BIGDEC
Each data type has several storage types each suited to optimize storage/retrieval depending on the cardinality and/or length of data. In AMIHDB, there are four storage types:
FLAT, BITMAP, VARSIZE, PARTITION.
Important Note: Different partitions can have different storage types for the same column.

VARSIZE

VARSIZE is the storage engine for variable-length data types:

STRING,BINARY,BIGINT,BIGDEC

There are three varieties of VARSIZE:

VARSIZE3,VARSIZE4,VARSIZE5 (default is VARSIZE5)

The number at the end indicates the number of bytes of the address pointer. For example, VARSIZE3 indicates that the pointer has an address size of 3 Bytes, meaning that it can address up to approximately 16 Megabytes of data (256*256*256).

Here are is the amount of total data that a column can hold depending on the type:

VARSIZE/Data Size
VARSIZE() Size
VARSIZE3 16MB
VARSIZE4 4GB
VARSIZE5 1TB

Note, that if data exceeds the size of a column type, AMIHDB will automatically reconfigure and extend the VARSIZE type (Note: this will be an expensive operation though). AMIHDB will default to VARSIZE5, if not explicitly specified in the CREATE TABLE schema.

The way AMIHDB stores a variable-length value is to keep an address pointer to point to the offset position relative to the first string to store each individual new string. As the diagram below shows, each time we are adding a new string, we are appending it to the end of the preceding string in the string file. In the example below, each number in the string file block indicates the number of bytes each string occupies. The first address pointer points to the first string of 17 Bytes long and since there is no preceding string, the offset value the pointer stores is 0. When we try to add the second string of length 23 Bytes long, we use the second address pointer, which points to the offset distance relative to the first string, which is 17, to point to the second string, so on and so forth.
MemoryPtr.png

As the diagram below shows, AMIHDB needs 2 files to store the String VARSIZE column. One file (<column_name>.3fhcol) is the array of offset values for each address pointer and the other file (<column_name>.3fhdat) stores the actual string payload.

MemoryPtr2.png.jpg

VARSIZE Example

Let’s create a historical table which contains VARSIZE storage types

CREATE PUBLIC table VARSExample(AllNames String, ShortNames String VARSIZE3, MediumNames String VARSIZE4, LongNames String VARSIZE5) USE PersistEngine="HISTORICAL";
Insert into VARSExample values("All the names","abc","abcdef","abcdefghijklmnopqrstuvwxyz");

If we diagnose the table, we can see detailed information for each of the string columns.

Diagnose table VARSExample;

VarsizeEG.png
We can see that the AMIHDB by default assigns the string variable to VARSIZE5. So if you are not declaring a column as " ColumnName String VARSIZE", AMIHDB will assume the worst case scenario by allocating the maximum memory to the string column (AKA VARSIZE5).

FLAT

FLAT is the storage type for data with a known fixed length. Data types that use FLAT are:
BYTE,SHORT,INT,LONG,FLOAT,DOUBLE,CHAR,UTC,UTCN,UUID,COMPLEX,BOOLEAN
There are three varieties of FLAT:
FLAT, FLAT_NONULL, FLAT_NOMIN (default is FLAT)

  • FLAT stores all values and null but uses an extra byte.
  • FLAT_NONULL can store all values except null.
  • FLAT_NOMIN cat store all values except min value and null
  • Note, that if an entry is inserted that does not fit in the column type (for example, inserting null

into a FLAT_NONULL column) AMIHDB will automatically reconfigure and extend the type so the entry will fit. AMIHDB will default to FLAT if not explicitly specified in the CREATE TABLE schema.

Here is an example of storage cost per row for INTEGER, which is 4 bytes:

  • INTEGER FLAT: 5 bytes in total, which consists of 4 bytes of data + 1 byte for NULL flag.
  • INTEGER FLAT_NONULL: 4 bytes in total (since there is no NULL flag to store)
  • INTEGER FLAT_NOMIN: 4 bytes of data in total (the minimum value of that data type is

reserved as a placeholder for null hence its not available)

FLAT Example

First let’s create a historical table with the different FLAT types and insert some rows into it. If we diagnose the table, we can see that once again, the AMIHDB is assuming the worst case scenario for the Integer column(FLAT) if you don’t declare the storage type for this column.

CREATE PUBLIC table FlatExample(id int, id2 int FLAT,id3 int FLAT_NOMIN, id4 int FLAT_NONULL) USE PersistEngine="HISTORICAL";
Insert into FlatExample values(1,23,NULL,-2147483648);
Diagnose table FlatExample;

FLAT eg.png

BITMAP

A BITMAP is ideal if the cardinality of the data is relatively low (aka the same values are recurring a lot in the table). The ONLY datatype that supports BITMAP is STRING.
There are 2 options for bitmap:
BITMAP1, BITMAP2 (BITMAP2 is the default)

The number indicates how many bytes are used to identify unique values. For example, a BITMAP1 stores 255 unique values (and NULL) while a BITMAP2 stores 65,535 unique values (and NULL). If the row has more than 65,535 unique entries, then a VARSIZE storage must be used instead.

BITMAP Example

First let’s create a historical table with a String BITMAP column and insert some rows into it.

CREATE PUBLIC table BITMAPExample(symbol1 String BITMAP1,symbol2 String BITMAP2) USE PersistEngine="HISTORICAL";
Insert into BITMAPExample values("IBM","AAPL"),("MSFT","TSLA"),("MSFT","FaceBook");
Diagnose table BITMAPExample;}

Bitmap eg.png

PARTITION

PARTITION is a special type of column that needs to be specified by the user in the CREATE TABLE statement. A partition column aggregates the rows with the same value into a single partition. For example, if the user defines a ”Date” column as a LONG PARTITION, then all the rows with the same date value will go into the same partition(essentially stored in the same file path). One can add more PARTITION columns to the table as needed. For example the user can declare two PARTITION columns:Date LONG and Region STRING. In this case, the rows with the same Date and the same Region name will go into the same partition and get stored in the same file path.

Important Note: It is strongly recommended to have at least one Partition column otherwise all data for the entire table will be stored in a single partition

Important Note: All partition columns must be declared up front, prior to the insertion of the data and is immutable. Immutability means that: (a). Once the column has been declared as a partition column, it cannot be altered afterwards.
(b). Vice versa if the column has been initially declared as a non-partition column, you cannot convert it into the partition column afterwards.
(c). The data in the PARTITION column cannot be updated/deleted

PARTITION Example

To illustrate how the data records are split across different partitions based on a PARTITION column, the following table PartitionDemo(date LONG PARTITION, symbol STRING, qty INTEGER, comments STRING) is constructed. "PARTITION" keyword is the way we declare a column as PARTITION. In simple terms, the rows with the same column value will be placed into the same directory if we declare a particular column to be PARTITION. The details of different storage types will be elaborated in the section “Data types and storage types”.

create public table PartitionDemo(date LONG PARTITION, symbol STRING, qty INTEGER, comments STRING) use PersistEngine="HISTORICAL";

If there exists a partition storage type in the table, like the example shown below where date is a LONG PARTITION column, then the rows with the same partition value will be placed into the same AMI History Partition, uniquely identified by a Partition Index(PID). For example if we have three data records:

(20161120,"IBM",30,"very good products"),
(20161122,"MSFT",60,"Nice shopping experience")
(20161122,"MSFT",60,"Very positive customer feedback")

insert into PartitionDemo values(20161120,"IBM",30,"very good products");
insert into PartitionDemo values(20161122,"MSFT",60,"Nice shopping experience");
insert into PartitionDemo values(20161122,"MSFT",60,"Very positive customer feedback");

Since “date” is declared as a LONG PARTITION column, the rows with the same “date” value will get clustered into one partition. In the case above, The first row and the second row with the same “date” value of 20161122 will be placed into the same partition. The third row will be placed into a separate “AMI History Partition”. If we enter the command “Diagnose table”, we can see detailed information about the AMI History Table.

Diagnose table PartitionDemo;

Partition eg1.png

We can see that there are two “AMI History Partitions” created, each uniquely identified by the Partition index. In the table diagnosis diagram above, we have the first “AMI History Partition”, uniquely identified by PID=1, and second uniquely identified by PID=2.
Note that all the rows are currently in append mode, meaning the rows can still be modified. But once we commit the data, those become read-only (AKA historical).
Partition eg1 illustrative.png

Once committed (let’s say we have just restarted the AMI ), in the comments section, the IN_APPEND_MODE and NOT_OPTIMIZED keywords will disappear.
Partition readonly.png

Now lets insert some more rows into the table and check the table status again.<br>
insert into PartitionDemo values("20161120","AAPL",40,"big blockbuster");
insert into PartitionDemo values("20161124","IBM",100,"extremely user-friendly");
insert into PartitionDemo values("20161124","MSFT",50,"enjoyable experience");
Diagnose table PartitionDemo;

Partition explanation2.png

Although the first row in the second insertion has exactly the same date value ”20161120” as the value in the AMI History Partition 1, AMIHDB will create a separate partition to store these rows (in this case stored in AMI History Partition 3, uniquely identified by PID=3).

Now if you commit the data (either by restarting AMI or calling Optimization procedure), the data will be written to disk and become historical. AMIHDB will check if the rows you want to commit contain the same partition value in the historical partitions and collapse them into one partition accordingly.
Partitions collapse.png

After the second commit, since partition 3 and partition 1 have the same partition value of “20161120”, we can see that they collapse into one partition, (AKA partition 3 has been merged into partition 1 and thereby increasing the COUNT by 1). Every time we commit the partition, in the “COMMENTS” field, they will appear as “HISTORICAL”, meaning that all the data has been properly written to disk and become read-only.

PARTITION Summary

To summarize:
(a). Each AMI History Partition is uniquely identified by one AMI History Partition Index, also known as PID.
(b). Historical data and data to be committed will live in separate AMI History Partitions.
If the historical table contains a PARTITION column:

(a). Rows with different PARTITION values will be placed into different AMI History Partitions.
(b). Before the commit, the new rows with the same PARTITION value as the historical rows will be placed in a separate AMI History Partition.
(c). Once we commit the data, AMI History Partitions with the same PARTITION value will collapse into one single AMI History Partition.
If the historical table does not contain a PARTITION column:
(a). All historical rows will live in one AMI History Partition and all rows to be committed will live in another AMI History Partition. In other words, there can be at most 2 AMI History Partitions.

Choosing the Best Column Type For Your Variable

There are some general rules of thumb for choosing your optimal storage type for each column based on the nature of your data. The detailed guidelines for a string variable and an integer variable are given as below. These are explained further in the section “Optimization of existing data”.AMIHDB will dynamically adjust the storage type for you to make sure your existing data follows the best practices and assigns the optimal storage type to each of the columns.

Choosing the Best Column Type For a String Variable

Below is a general comparison of different VARSIZE options when declaring a string column.

String Variable Storage Options
Type Storage Method Address pointer size Max File Size Advantage Disadvantage Usage Scenario
String VARSIZE3 Each address pointer has 3 bytes of storage, resulting in max file size of 16777216 Bytes 3 Bytes = 24 Bits Maximum file size = 2^24 Bytes = 16777216 Bytes = 16MB Takes up least amount of storage Cannot store strings over 16MB total size To store strings not too long in total size
String VARSIZE4 Each address pointer has 4 bytes of storage, resulting in max file size of 4096 MegaBytes 4 Bytes = 32 Bits Maximum file size = 2^32 Bytes = 4294967296 Bytes = 4096MB Takes up medium amount of storage Cannot store strings over 4096MB total size. There is some memory overhead To store strings whose total size does not exceed 4096MB
String VARSIZE5 Each address pointer has 5 bytes of storage, resulting in max file size of 1024 GigaBytes 5 Bytes = 40 Bits Maximum file size = 2^40 Bytes = 1099511600000 Bytes = 1024GB Can store a huge amount of strings up to 1024GB Takes up most amount of storage and therefore a lot of memory overhead To store strings that can be astronomically long in size(up tp 1TB in size)
String (By Default) Assume the worst case scenario, allocating the maximum space to each address pointer (i.e.VARSIZE5) 5 Bytes = 40 Bits Maximum file size = 2^40 Bytes = 1099511600000 Bytes = 1024GB Can store a huge amount of strings up to 1024GB Takes up most amount of storage and therefore a lot of memory overhead To store strings that can be astronomically long in size(up tp 1TB in size)
String BITMAP1 If the number of unique strings is between 1-255 NOT relevant NOT relevant Very memory efficient and fastest retrieval NAN For example we only have 10 unique strings out of a large number of rows
String BITMAP2 If the number of unique strings is between 255-65,535 NOT relevant NOT relevant Very memory efficient and fastest retrieval NAN For example we have 10,000 unique values in the column

Choosing the Best Column Type For an Integer Variable

The similar principles can also be applied to BYTE, SHORT, LONG and BOOLEAN columns that have fixed size.

Integer Variable Storage Options
Type Storage Advantage Disadvantage
Integer FLAT Can store NULL+MINVALUE Big memory space. Can store both NULL and MINVALUE Can be memory inefficient if the actual data set does not actually need this much memory space.
Integer FLAT_NOMIN Can store NULL Can store NULL flag Cannot store MINVALUE
Integer FLAT_NONULL Can store MINVALUE Can store MINVALUE Cannot store NULL flag
Integer (By default) Default setting for Integer. Assume worst case scenario, same as "Integer FLAT" Big memory space. Can store both NULL and MINVALUE Can be memory inefficient if the actual data set does not need both min and null values.

Advanced

This section will go more in-depth into how AMI historical database optimizes itself for more efficient data storage, how the table schema can be altered, how the index can be created as well as how to perform update/delete operations.

Optimization on Existing Data

New partitions are, by default, optimized for flexible insertion of data. After all data has been added for a partition, the partition can be “Optimized” which will reorganize the partition data to improve read performance and minimize disk storage. Additionally, the table’s schema can provide “hints” on how to optimize columns ahead of time, avoiding the cost of optimization later.

The AMIHDB dynamically adjusts the storage type according to how much space your existing data needs for each of your data columns. One could specify the schema of the table at the very beginning, for example we have a column manually specified as quantity Integer FLAT_NONULL. Later on while we are inserting data, if we suddenly need to insert a NULL flag to the table, although the schema says FLAT_NONULL, the AMIHDB is smart enough to alter the schema to FLAT_NOMIN so that we can store the NULL flag in the table. There are other optimization examples that will elaborate how optimization works in various scenarios.

FLAT Optimization Example

First let’s create a historical table with an integer id column.

CREATE PUBLIC TABLE test(id int) USE PersistEngine="HISTORICAL";
Insert into test values(1),(2),(3);
  • By default, because we did not specify a storage type before we optimized the table, AMIHDB

will assume the worst case scenario for data storage and assign FLAT to the integer column. If we diagnose the table, we will see that their id storage type is Flat.

FLAT DEFAULT.png

  • Next, let’s optimize the table and DIAGNOSE TABLE again to see how the storage type for the

id column will change. We optimize the historical table by calling the __OPTIMIZE_HISTORICAL_TABLE(TableName) stored procedure. We see the storage type for the id column has changed to FLAT_NOMIN since AMIHDB scans and confirms that a minimum value to store does not exist.

CALL __OPTIMIZE_HISTORICAL_TABLE("test");
DIAGNOSE TABLE test;

FLAT postOptimize.png

  • Then let’s truncate the table and try inserting the minimum value. Let’s see how the storage type

changes accordingly. Since the minimum value -2147483648 exists, which is the most negative value Java can store, the storage type will change to FLAT_NONULL to adapt to the data size in the table.

TRUNCATE TABLE test;
INSERT INTO test VALUES(1),(-2147483648);
CALL __OPTIMIZE_HISTORICAL_TABLE("test");
DIAGNOSE TABLE test;

FLAT NONULL WHEN MINinserted.png

  • Finally, let’s truncate the table again and try inserting the NULL value. Since a NULL value

exists, the storage type is optimized to FLAT_NOMIN.

TRUNCATE TABLE test;
INSERT INTO test VALUES(1),(NULL);
CALL __OPTIMIZE_HISTORICAL_TABLE("test");
DIAGNOSE TABLE test;

FLAT postOptimize.png

VARSIZE Optimization Example

  • First let’s create a testing table with a string column. Before optimization, once again the

AMIHDB will assume the greatest storage type requirement by assigning VARSIZE5 to the string variable

CREATE PUBLIC TABLE varTest(name STRING) USE PersistEngine="HISTORICAL";
INSERT INTO varTest VALUES("test1"),("test2"),("test3");
DIAGNOSE TABLE varTest;

VARSIZE5.png

  • Now, let’s optimize the table. We can see that the storage type for the string variable name

has been changed to VARSIZE3 because the database realizes that VARSIZE3 is sufficient to store all the strings test1, test2, test3

CALL __OPTIMIZE_HISTORICAL_TABLE("varTest");
DIAGNOSE TABLE varTest;

VARSIZE3.png

  • Next, let’s insert some more strings into the table and optimize it to see how the storage

type will change. As we insert 1 million rows into the table and do the optimization, we can see that the storage type has been changed from VARSIZE3 to VARSIZE4.

for (int i =0;i<1000000;i++) INSERT INTO varTest
VALUES("test"+rand());
DIAGNOSE TABLE varTest;

VARSIZE4.png

BITMAP Optimization Example

  • Let’s create a table with a string column and insert 10,000 rows into the table. Before we do the

optimization, we can see the storage type for the string is VARSIZE5

CREATE PUBLIC TABLE BmapTest(name string) USE PersistEngine="HISTORICAL";
For (int i =0;i<10000;i++) {INSERT INTO BmapTest VALUES("test"+rand(10));}
DIAGNOSE TABLE BmapTest;


BITMAP VARSIZE5.png

  • After the optimization, the AMIHDB realizes that there are only 10 unique values in the table for

the string and adjusts the storage type to BITMAP1 accordingly.

CALL __OPTIMIZE_HISTORICAL_TABLE("BmapTest");
DIAGNOSE TABLE BmapTest;

PARTITION Optimization Example

  • Now let’s create a historical table with one column as a string partition and another column as

an integer. We can see that the rows with the same value are placed into the same partition. All rows with the string value equal to test1 are placed into one partition and all rows with the string value equal to test2 are placed into the other partition.

CREATE PUBLIC TABLE PTest(name string partition, qty int) USE PersistEngine="HISTORICAL";
INSERT INTO PTest VALUES("test1",20),("test1",30),("test1",40),("test2",20),("test2",120);
DIAGNOSE TABLE PTest;

PARITION OPTIMIZE.png

  • If we do the optimization on the table, all the rows that we inserted will get optimized and be in

read-only mode. This is confirmed by the lack of IN_APPEND_MODE and NOT_OPTIMIZED in the comments section. This means that all data records are properly written to disk and marked as HISTORICAL. The optimization automatically happens after the session has been closed, which results in all data writing to disk.

CALL __OPTIMIZE_HISTORICAL_TABLE("PTest");
DIAGNOSE TABLE PTest;

PTest1.png

  • If the user wants to insert more data into the historical database afterwards in a new session

(Here calling optimization procedure is equivalent to starting a new session). We can see from diagnosing the table that, instead of appending the new data records into the historical partition columns, we create brand new partitions for them.

INSERT INTO PTest VALUES("test1",60),("test2",35),("test2",25),("test2",10);
DIAGNOSE TABLE PTest;

PTest2.png

  • Now if we call the optimization procedure against the historical table, the AMIHDB will look to

see if there are partitions with the same partition keys that can be merged with the historical partitions and will do so if there are. Here we can see the new partition with the partition key of test1 and test2 get merged with the historical partitions with the same keys.

CALL __OPTIMIZE_HISTORICAL_TABLE("PTest");
DIAGNOSE TABLE PTest;

PTest3.png

File Directory

Back to the very first example where we explained how data gets partitioned across different columns, we used the table PartitionDemo(date LONG PARTITION, symbol STRING, qty INTEGER, comments STRING).
In this section, from a file directory perspective, we will take a closer look at how data in each column is stored across different files.

CREATE PUBLIC TABLE PartitionDemo(date LONG PARTITION, symbol STRING, qty INTEGER, comments STRING) USE PersistEngine="HISTORICAL";

In the table PartitionDemo,we have these rows:
Partitiondemo Data.png
And we have 3 Partitions, each uniquely identified by a PID.

FileDiretory2.png

If we go to /ami/amione/hdb directory, we will see all the folders, named as Your_Historical_Table_Name.3fhtab , where all the rows from the table are stored.
FolderDir.png