Ibis
Explore large datasets in HEAVY.AI with the full power of SQL, with a pandas-like API
Ibis allows you to explore data in a variety of remote storage systems, with an API is inspired by the pandas toolkit.
You can use Ibis to interact directly with HeavyDB and several other supported SQL systems by writing high-level Python code instead of lower-level SQL. Using familiar tools can increase productivity.
Ibis support for several execution backends including common SQL databases and processing engines. This allows you to author analytical workflows or operations once and run them consistently against these backends. You can also create complex workflows that can simultaneously materialize data from multiple backends and combine the resulting outputs in various ways.
Supported backends include PySpark, Google BigQuery, and PostgreSQL.
You can install ibis-heavyai using
conda (recommended) or mamba, which is a faster implementation of conda. For more information, see the Mamba documentation.conda install -c conda-forge ibis-heavyai
# or to install rbc to a new environemnt, run
conda create -n ibis -c conda-forge ibis-heavyai
conda activate ibis
# check if ibis-heavyai installed succesfully
python -c 'import ibis; ibis.heavyai.version'
You can also use
pip for package management:pip install ibis-heavyai
# check if ibis-heavyai installed succesfully
python -c 'import ibis; ibis.heavyai.version'
This short example demonstrates Ibis. Inside a notebook cell, you can first connect to an HeavyDB database instance:
import ibis
con = ibis.heavyai.connect(
host='localhost',
database='heavyai',
user='admin',
password='HyperInteractive'
)
If you are launching JupyterLab from Immerse, this connection is already set up for you based on your Immerse login and credentials.
This connection object has several utility methods that can come in handy, for listing database objects, creating tables, loading data etc.
Next, let's identify a table and define a simple Ibis expression against it:
tablename = 'github'
t = con.table(tablename)
count_expr = t.count()
count_expr.compile()
This expression is compiled by Ibis into SQL that runs against HeavyDB:
SELECT count(*) AS "count"
FROM github
When executed, the above SQL, as expected, counts the rows of the table:
print(f'Count of records in table {tablename} is {count_expr.execute()}')
Count of records in table github is 2033049065
You can also write raw SQL in Ibis. A SQL query can be wrapped as an Ibis expression and composed together with other Ibis expressions.
Ibis offers a powerful way to compose sophisticated analytical expressions, and then leverage the processing power and scale of supported storage backends to evaluate them and return results. The following shows a more complex Ibis expression and the SQL it generates. You can leverage a full programming language (Python) to assemble complex expressions that cover SQL features like joins, filters, and aggregates.
well_list = ['4201334812', '4201335379', '4201335112', '4201334947', '4201334329', '4201334965', '4201334978', '4201334998', '4201335467', '4201334917', '4201334930', '4212331826', '4212331164', '4212300017', '4212331782', '4212332076', '4212331839', '4228532777', '4228533451', '4212332354', '4212332207', '4212330405', '4212332014']
#Find oil well production for the specific well list above
sum_oil=t.oil_monthly.sum().log10().name('oil_monthly')
expr = (t.filter([t.api_no_10.isin(well_list), t.oil_monthly.notnull()])\
.group_by(['date_prod', 'api_no_10']).having(sum_oil >= 0.)\
.aggregate(sum_oil))\
.sort_by('date_prod')
# find the max production month per well in the wellset
max_oil = expr.oil_monthly.max().name('oil_monthly')
expr1 = expr.group_by('api_no_10').aggregate(max_oil)
expr2 = (expr.join(expr1, [expr.api_no_10 == expr1.api_no_10,\
expr.oil_monthly == expr1.oil_monthly]))\
[expr.api_no_10.name("api"), expr.oil_monthly.name("om2"), expr.date_prod.name("dp2")]
print(expr2.compile())
This results in automatically generated SQL:
WITH t0 AS (
SELECT "date_prod", "api_no_10", log10(sum("oil_monthly")) AS oil_monthly
FROM upstream_reservoir
WHERE ("api_no_10" IN ('4201334998', '4212331782', '4212331826', '4201334965', '4212300017', '4201334329', '4228533451', '4212331164', '4201335467', '4201334930', '4212330405', '4201334947', '4212332354', '4201334978', '4201334917', '4201335379', '4201334812', '4212332207', '4228532777', '4201335112', '4212332076', '4212331839', '4212332014')) AND
("oil_monthly" IS NOT NULL)
GROUP BY date_prod, api_no_10
HAVING log10(sum("oil_monthly")) >= 0.0
ORDER BY "date_prod"
)
SELECT t0."api_no_10" AS api, t0."oil_monthly" AS om2, t0."date_prod" AS dp2
FROM t0
JOIN (
SELECT "api_no_10", max("oil_monthly") AS oil_monthly
FROM t0
GROUP BY api_no_10
) t1
ON (t0."api_no_10" = t1."api_no_10") AND
(t0."oil_monthly" = t1."oil_monthly")
The result of the evaluation is by default a pandas dataframe, making it convenient to use Ibis inside other Python data analysis workflows.
expr2.execute()
| api | om2 | dp2 |
0 | 4201335379 | 3.944927 | 2016-11-01 |
1 | 4201335467 | 4.125253 | 2017-12-01 |
2 | 4201334998 | 3.822887 | 2014-09-01 |
3 | 4201334930 | 3.698709 | 2014-05-01 |
4 | 4201334917 | 4.353243 | 2014-06-01 |
5 | 4201334812 | 4.127299 | 2014-03-01 |
6 | 4201334947 | 3.715502 | 2014-06-01 |
7 | 4201334978 | 3.822887 | 2014-09-01 |
8 | 4201335112 | 4.010766 | 2015-01-01 |
9 | 4201334329 | 3.721068 | 2015-07-01 |
Although Ibis can output to pandas, the result is first materialized and could be transferred over the network to the client running Ibis (such as a Jupyter notebook in a browser on a laptop). Avoid using it for expressions that try to move a million rows to a browser.
Ibis depends on heavyai as a low-level connector. It can leverage some of the key features of pyomnisci, including output of query results to the Arrow format via the
select_ipcandselect_ipc_gpucalls.Building on this capability, the HeavyDB backend for Ibis can output query results using Apache Arrow to CPU or GPU memory (the latter by using the cudf dataframe library from NVIDIA). This makes it convenient to use with the GPU-based RAPIDS machine learning methods. Output to cudf is as simple as passing an extra couple of parameters to the
execute method in ibis, for example, as followsgdf = get_ts_by_period_station(zip_cd=94568).execute(gpu_device=3, ipc=True)
RAPIDs is currently usable only with supported Nvidia GPUs and requires CUDA
The HeavyDB backend for Ibis supports geospatial functions building directly on HeavyDB SQL support for these functions. These operations produce geopandas dataframes, allowing for usage in Python-based geospatial analytics workflows.
Here is an example of how a table that has geospatial data in HEAVY.AI, can output directly to a geopandas dataframe.
heavyai_cli = ibis.heavyai.connect(
host='metis.heavyai.com', user='heavyai', password='HyperInteractive',
port=443, database='heavyai', protocol= 'https'
)
t = heavyai_cli.table('zipcodes_2017')
df = t.head().execute()
type(df)
#geopandas.geodataframe.GeoDataFrame
Using Ibis, you can create and use user-defined functions (UDFs) in Python that execute inside HeavyDB. This UDF framework leverages integration with Numba, a Just-in-Time (JIT) compiler backend for python, and produces lower-level code for a more performant execution path than within Python itself.
It also makes it easy to author UDFs in Python, and then make the UDFs usable in a SQL workflow.
See the following Ibis documentation to get started:
- The Ibis User Guide---in particular, the Ibis for SQL programmers section, which shows how Ibis simplifies workflows involving analytical SQL.