Develop Reference

Py-Spark mapPartitions: how to craft the function?

We are using Databricks on Azure with a reasonably large cluster (20 cores, 70GB memory across 5 executors). I have a parquet file with 4 million rows. Spark can read well, call that sdf.
I am hitting the problem that the data must be converted to a Pandas dataframe. Taking the easy/obvious way pdf = sdf.toPandas() causes an out of memory error.
So I want to apply my function separately to subsets of the Spark DataFrame. The sdf itself is in 19 partitions, so what I want to do is write a function and apply it to each partition separately. Here's where mapPartitions comes in.
I was trying to write my own function like
def example_function(sdf):
pdf = sdf.toPandas()
/* apply some Pandas and Python functions we've written to handle pdf.*/
output = great_function(pdf)
return output
Then I'd use mapPartitions to run that.
sdf.rdd.mapPartitions(example_function)
That fails with all kinds of errors.
Looking back at the instructions, I realize I'm clueless! Iwas too optimistic/simplistic in what they expect to get from me. They don't seem to imagine that I'm using my own functions to handle the whole Spark DF that exists partition. They seem to plan only for code that would handle the rows in the Spark data frame one row at a time and the parameters are Iterators.
Can you please share you thoughts on this?

In your example case it might be counter productive to start from a Spark Dataframe and fall back to RDD if you're aiming at using pandas.
Under the hood toPandas() is triggering collect() which retrieve all data on the driver node, which will fail on large data.
If you want to use pandas code on Spark, you can use pandas UDFs which are equivalent to UDFs but designed and optimized for pandas code.
https://docs.databricks.com/spark/latest/spark-sql/udf-python-pandas.html

I did not find a solution using Spark map or similar. Here is best option I've found.
The parquet folder has lots of smaller parquet files inside it. As long as default settings were used, these files have extension snappy.parquet. Use Python os.listdir and filter out the file list to ones with correct extension.
Use Python and Pandas, NOT SPARK, tools to read the individual parquet files. It is much faster to load a parquet file with a few 100,000 rows with pandas than it is with Spark.
For the loaded dataframes, run the function I described in the first message, where the dataframe gets put through the wringer.
def example_function(pdf):
/* apply some Pandas and Python functions we've written to handle pdf.*/
output = great_function(pdf)
return output
Since the work for each data section has to happen in Pandas anyway, there's no need to keep fighting with Spark tools.
Other bit worth mentioning is that joblib's Parallel tool can be used to distribute this work among cluster nodes.

dask out of memory error despite data size doesnt exceed memory

I'm trying to load a dask dataframe from a MySQL table which takes about 4gb space on disk. I'm using a single machine with 8gb of memory but as soon as I do a drop duplicate and try to get the length of the dataframe, an out of memory error is encountered.
Here's a snippet of my code:
df = dd.read_sql_table("testtable", db_uri, npartitions=8, index_col=sql.func.abs(sql.column("id")).label("abs(id)"))
df = df[['gene_id', 'genome_id']].drop_duplicates()
print(len(df))
I have tried more partitions for the dataframe(as many as 64) but they also failed. I'm confused why this could cause an OOM? The dataframe should fit in memory even without any parallel processing.

which takes about 4gb space on disk
It is very likely to be much much bigger than this in memory. Disk storage is optimised for compactness, with various encoding and compression mechanisms.
The dataframe should fit in memory
So, have you measured its size as a single pandas dataframe?
You should also keep in mind than any processing you do to your data often involves making temporary copies within functions. For example, you can only drop duplicates by first finding duplicates, which must happen before you can discard any data.
Finally, in a parallel framework like dask, there may be multiple threads and processes (you don't specify how you are running dask) which need to marshal their work and assemble the final output while the client and scheduler also take up some memory. In short, you need to measure your situation, perhaps tweak worker config options.

You don't want to read an entire DataFrame into a Dask DataFrame and then perform filtering in Dask. It's better to perform filtering at the database level and then read a small subset of the data into a Dask DataFrame.
MySQL can select columns and drop duplicates with distinct. The resulting data is what you should read in the Dask DataFrame.
See here for more information on syntax. It's easiest to query databases that have official connectors, like dask-snowflake.

Efficient use of dask with parquet files

I have received a huge (140MM records) dataset and Dask has come in handy but I'm not sure if I could perhaps do a better job. Imagine the records are mostly numeric (two columns are dates), so the process to transform from CSV to parquet was a breeze (dask.dataframe.read_csv('in.csv').to_parquet('out.pq')), but
(i) I would like to use the data on Amazon Athena, so a single parquet file would be nice. How to achieve this? As it stands, Dask saved it as hundreds of files.
(ii) For the Exploratory Data Analysis I'm trying with this dataset, there are certain operations where I need more then a couple of variables, which won't fit into memory so I'm constantly dumping two/three-variable views into SQL, is this code efficient use of dask?
mmm = ['min','mean','max']
MY_COLUMNS = ['emisor','receptor','actividad', 'monto','grupo']
gdict = {'grupo': mmm, 'monto': mmm, 'actividad': ['mean','count']}
df = dd.read_parquet('out.pq', columns=MY_COLUMNS).groupby(['emisor','receptor']).agg(gdict)
df = df.compute()
df.columns = ['_'.join(c) for c in df.columns] # ('grupo','max') -> grupo_max
df.to_sql('er_stats',conn,index=False,if_exists='replace')
Reading the file takes about 80 and writing to SQL about 60 seconds.

To reduce the number of partitions, you should either set the blocksize when reading the CSV (preferred), or repartition before writing the parquet. The "best" size depends on your memory and number of workers, but a single partition is probably not possible if your data is "huge". Putting the many partitions into a single file is also not possible (or, rather, not implemented), because dask writes in parallel and there would be no way of knowing where in the file the next part goes before the previous part is finished. I could imagine writing code to read in successive dask-produced parts and streaming them into a single output, it would not be hard but perhaps not trivial either.
writing to SQL about 60 seconds
This suggests that your output is still quite large. Is SQL the best option here? Perhaps writing again to parquet files would be possible.

partitionBy taking too long while saving a dataset on S3 using Pyspark

I am trying to save a dataset using partitionBy on S3 using pyspark. I am partitioning by on a date column. Spark job is taking more than hour to execute it. If i run the code without partitionBy it just takes 3-4 mints.
Could somebody help me in fining tune the parititonby?

Ok, so spark is terrible at doing IO. Especially with respect to s3. Currently when you are writing in spark it will use a whole executor to write the data SEQUENTIALLY. That with the back and forth between s3 and spark leads to it being quite slow. So you can do a few things to help mitigate/side step these issues.
Use a different partitioning strategy, if possible, with the goal being minimizing files written.
If there is a shuffle involved before the write, you can change the settings around default shuffle size: spark.sql.shuffle.partitions 200 // 200 is the default you'll probably want to reduce this and/or repartition the data before writing.
You can go around sparks io and write your own hdfs writer or use s3 api directly. Using something like foreachpartition then a function for writing to s3. That way things will write in parallel instead of sequentially.
Finally, you may want to use repartition and partitionBy together when writing (DataFrame partitionBy to a single Parquet file (per partition)). This will lead to one file per partition when mixed with maxRecordsPerFile (below) above this will help keep your file size down.
As a side note: you can use the option spark.sql.files.maxRecordsPerFile 1000000 to help control file sizes to make sure they don't get out of control.
In short, you should avoid creating too many files, especially small ones. Also note: you will see a big performance hit when you go to read those 2000*n files back in as well.
We use all of the above strategies in different situations. But in general we just try to use a reasonable partitioning strategy + repartitioning before write. Another note: if a shuffle is performed your partitioning is destroyed and sparks automatic partitioning takes over. Hence, the need for the constant repartitioning.
Hope these suggestions help. SparkIO is quite frustrating but just remember to keep files read/written to a minimum and you should see fine performance.

Use version 2 of the FileOutputCommiter
.set("mapreduce.fileoutputcommitter.algorithm.version", "2")

"Large data" workflows using pandas [closed]

Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
We don’t allow questions seeking recommendations for books, tools, software libraries, and more. You can edit the question so it can be answered with facts and citations.
Closed 1 year ago.
The community reviewed whether to reopen this question 1 year ago and left it closed:
Original close reason(s) were not resolved
Improve this question
I have tried to puzzle out an answer to this question for many months while learning pandas. I use SAS for my day-to-day work and it is great for it's out-of-core support. However, SAS is horrible as a piece of software for numerous other reasons.
One day I hope to replace my use of SAS with python and pandas, but I currently lack an out-of-core workflow for large datasets. I'm not talking about "big data" that requires a distributed network, but rather files too large to fit in memory but small enough to fit on a hard-drive.
My first thought is to use HDFStore to hold large datasets on disk and pull only the pieces I need into dataframes for analysis. Others have mentioned MongoDB as an easier to use alternative. My question is this:
What are some best-practice workflows for accomplishing the following:
Loading flat files into a permanent, on-disk database structure
Querying that database to retrieve data to feed into a pandas data structure
Updating the database after manipulating pieces in pandas
Real-world examples would be much appreciated, especially from anyone who uses pandas on "large data".
Edit -- an example of how I would like this to work:
Iteratively import a large flat-file and store it in a permanent, on-disk database structure. These files are typically too large to fit in memory.
In order to use Pandas, I would like to read subsets of this data (usually just a few columns at a time) that can fit in memory.
I would create new columns by performing various operations on the selected columns.
I would then have to append these new columns into the database structure.
I am trying to find a best-practice way of performing these steps. Reading links about pandas and pytables it seems that appending a new column could be a problem.
Edit -- Responding to Jeff's questions specifically:
I am building consumer credit risk models. The kinds of data include phone, SSN and address characteristics; property values; derogatory information like criminal records, bankruptcies, etc... The datasets I use every day have nearly 1,000 to 2,000 fields on average of mixed data types: continuous, nominal and ordinal variables of both numeric and character data. I rarely append rows, but I do perform many operations that create new columns.
Typical operations involve combining several columns using conditional logic into a new, compound column. For example, if var1 > 2 then newvar = 'A' elif var2 = 4 then newvar = 'B'. The result of these operations is a new column for every record in my dataset.
Finally, I would like to append these new columns into the on-disk data structure. I would repeat step 2, exploring the data with crosstabs and descriptive statistics trying to find interesting, intuitive relationships to model.
A typical project file is usually about 1GB. Files are organized into such a manner where a row consists of a record of consumer data. Each row has the same number of columns for every record. This will always be the case.
It's pretty rare that I would subset by rows when creating a new column. However, it's pretty common for me to subset on rows when creating reports or generating descriptive statistics. For example, I might want to create a simple frequency for a specific line of business, say Retail credit cards. To do this, I would select only those records where the line of business = retail in addition to whichever columns I want to report on. When creating new columns, however, I would pull all rows of data and only the columns I need for the operations.
The modeling process requires that I analyze every column, look for interesting relationships with some outcome variable, and create new compound columns that describe those relationships. The columns that I explore are usually done in small sets. For example, I will focus on a set of say 20 columns just dealing with property values and observe how they relate to defaulting on a loan. Once those are explored and new columns are created, I then move on to another group of columns, say college education, and repeat the process. What I'm doing is creating candidate variables that explain the relationship between my data and some outcome. At the very end of this process, I apply some learning techniques that create an equation out of those compound columns.
It is rare that I would ever add rows to the dataset. I will nearly always be creating new columns (variables or features in statistics/machine learning parlance).

I routinely use tens of gigabytes of data in just this fashion
e.g. I have tables on disk that I read via queries, create data and append back.
It's worth reading the docs and late in this thread for several suggestions for how to store your data.
Details which will affect how you store your data, like:
Give as much detail as you can; and I can help you develop a structure.
Size of data, # of rows, columns, types of columns; are you appending
rows, or just columns?
What will typical operations look like. E.g. do a query on columns to select a bunch of rows and specific columns, then do an operation (in-memory), create new columns, save these.
(Giving a toy example could enable us to offer more specific recommendations.)
After that processing, then what do you do? Is step 2 ad hoc, or repeatable?
Input flat files: how many, rough total size in Gb. How are these organized e.g. by records? Does each one contains different fields, or do they have some records per file with all of the fields in each file?
Do you ever select subsets of rows (records) based on criteria (e.g. select the rows with field A > 5)? and then do something, or do you just select fields A, B, C with all of the records (and then do something)?
Do you 'work on' all of your columns (in groups), or are there a good proportion that you may only use for reports (e.g. you want to keep the data around, but don't need to pull in that column explicity until final results time)?
Solution
Ensure you have pandas at least 0.10.1 installed.
Read iterating files chunk-by-chunk and multiple table queries.
Since pytables is optimized to operate on row-wise (which is what you query on), we will create a table for each group of fields. This way it's easy to select a small group of fields (which will work with a big table, but it's more efficient to do it this way... I think I may be able to fix this limitation in the future... this is more intuitive anyhow):
(The following is pseudocode.)
import numpy as np
import pandas as pd
# create a store
store = pd.HDFStore('mystore.h5')
# this is the key to your storage:
# this maps your fields to a specific group, and defines
# what you want to have as data_columns.
# you might want to create a nice class wrapping this
# (as you will want to have this map and its inversion)
group_map = dict(
A = dict(fields = ['field_1','field_2',.....], dc = ['field_1',....,'field_5']),
B = dict(fields = ['field_10',...... ], dc = ['field_10']),
.....
REPORTING_ONLY = dict(fields = ['field_1000','field_1001',...], dc = []),
)
group_map_inverted = dict()
for g, v in group_map.items():
group_map_inverted.update(dict([ (f,g) for f in v['fields'] ]))
Reading in the files and creating the storage (essentially doing what append_to_multiple does):
for f in files:
# read in the file, additional options may be necessary here
# the chunksize is not strictly necessary, you may be able to slurp each
# file into memory in which case just eliminate this part of the loop
# (you can also change chunksize if necessary)
for chunk in pd.read_table(f, chunksize=50000):
# we are going to append to each table by group
# we are not going to create indexes at this time
# but we *ARE* going to create (some) data_columns
# figure out the field groupings
for g, v in group_map.items():
# create the frame for this group
frame = chunk.reindex(columns = v['fields'], copy = False)
# append it
store.append(g, frame, index=False, data_columns = v['dc'])
Now you have all of the tables in the file (actually you could store them in separate files if you wish, you would prob have to add the filename to the group_map, but probably this isn't necessary).
This is how you get columns and create new ones:
frame = store.select(group_that_I_want)
# you can optionally specify:
# columns = a list of the columns IN THAT GROUP (if you wanted to
# select only say 3 out of the 20 columns in this sub-table)
# and a where clause if you want a subset of the rows
# do calculations on this frame
new_frame = cool_function_on_frame(frame)
# to 'add columns', create a new group (you probably want to
# limit the columns in this new_group to be only NEW ones
# (e.g. so you don't overlap from the other tables)
# add this info to the group_map
store.append(new_group, new_frame.reindex(columns = new_columns_created, copy = False), data_columns = new_columns_created)
When you are ready for post_processing:
# This may be a bit tricky; and depends what you are actually doing.
# I may need to modify this function to be a bit more general:
report_data = store.select_as_multiple([groups_1,groups_2,.....], where =['field_1>0', 'field_1000=foo'], selector = group_1)
About data_columns, you don't actually need to define ANY data_columns; they allow you to sub-select rows based on the column. E.g. something like:
store.select(group, where = ['field_1000=foo', 'field_1001>0'])
They may be most interesting to you in the final report generation stage (essentially a data column is segregated from other columns, which might impact efficiency somewhat if you define a lot).
You also might want to:
create a function which takes a list of fields, looks up the groups in the groups_map, then selects these and concatenates the results so you get the resulting frame (this is essentially what select_as_multiple does). This way the structure would be pretty transparent to you.
indexes on certain data columns (makes row-subsetting much faster).
enable compression.
Let me know when you have questions!

I think the answers above are missing a simple approach that I've found very useful.
When I have a file that is too large to load in memory, I break up the file into multiple smaller files (either by row or cols)
Example: In case of 30 days worth of trading data of ~30GB size, I break it into a file per day of ~1GB size. I subsequently process each file separately and aggregate results at the end
One of the biggest advantages is that it allows parallel processing of the files (either multiple threads or processes)
The other advantage is that file manipulation (like adding/removing dates in the example) can be accomplished by regular shell commands, which is not be possible in more advanced/complicated file formats
This approach doesn't cover all scenarios, but is very useful in a lot of them

There is now, two years after the question, an 'out-of-core' pandas equivalent: dask. It is excellent! Though it does not support all of pandas functionality, you can get really far with it. Update: in the past two years it has been consistently maintained and there is substantial user community working with Dask.
And now, four years after the question, there is another high-performance 'out-of-core' pandas equivalent in Vaex. It "uses memory mapping, zero memory copy policy and lazy computations for best performance (no memory wasted)." It can handle data sets of billions of rows and does not store them into memory (making it even possible to do analysis on suboptimal hardware).

If your datasets are between 1 and 20GB, you should get a workstation with 48GB of RAM. Then Pandas can hold the entire dataset in RAM. I know its not the answer you're looking for here, but doing scientific computing on a notebook with 4GB of RAM isn't reasonable.

I know this is an old thread but I think the Blaze library is worth checking out. It's built for these types of situations.
From the docs:
Blaze extends the usability of NumPy and Pandas to distributed and out-of-core computing. Blaze provides an interface similar to that of the NumPy ND-Array or Pandas DataFrame but maps these familiar interfaces onto a variety of other computational engines like Postgres or Spark.
Edit: By the way, it's supported by ContinuumIO and Travis Oliphant, author of NumPy.

This is the case for pymongo. I have also prototyped using sql server, sqlite, HDF, ORM (SQLAlchemy) in python. First and foremost pymongo is a document based DB, so each person would be a document (dict of attributes). Many people form a collection and you can have many collections (people, stock market, income).
pd.dateframe -> pymongo Note: I use the chunksize in read_csv to keep it to 5 to 10k records(pymongo drops the socket if larger)
aCollection.insert((a[1].to_dict() for a in df.iterrows()))
querying: gt = greater than...
pd.DataFrame(list(mongoCollection.find({'anAttribute':{'$gt':2887000, '$lt':2889000}})))
.find() returns an iterator so I commonly use ichunked to chop into smaller iterators.
How about a join since I normally get 10 data sources to paste together:
aJoinDF = pandas.DataFrame(list(mongoCollection.find({'anAttribute':{'$in':Att_Keys}})))
then (in my case sometimes I have to agg on aJoinDF first before its "mergeable".)
df = pandas.merge(df, aJoinDF, on=aKey, how='left')
And you can then write the new info to your main collection via the update method below. (logical collection vs physical datasources).
collection.update({primarykey:foo},{key:change})
On smaller lookups, just denormalize. For example, you have code in the document and you just add the field code text and do a dict lookup as you create documents.
Now you have a nice dataset based around a person, you can unleash your logic on each case and make more attributes. Finally you can read into pandas your 3 to memory max key indicators and do pivots/agg/data exploration. This works for me for 3 million records with numbers/big text/categories/codes/floats/...
You can also use the two methods built into MongoDB (MapReduce and aggregate framework). See here for more info about the aggregate framework, as it seems to be easier than MapReduce and looks handy for quick aggregate work. Notice I didn't need to define my fields or relations, and I can add items to a document. At the current state of the rapidly changing numpy, pandas, python toolset, MongoDB helps me just get to work :)

One trick I found helpful for large data use cases is to reduce the volume of the data by reducing float precision to 32-bit. It's not applicable in all cases, but in many applications 64-bit precision is overkill and the 2x memory savings are worth it. To make an obvious point even more obvious:
>>> df = pd.DataFrame(np.random.randn(int(1e8), 5))
>>> df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 100000000 entries, 0 to 99999999
Data columns (total 5 columns):
...
dtypes: float64(5)
memory usage: 3.7 GB
>>> df.astype(np.float32).info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 100000000 entries, 0 to 99999999
Data columns (total 5 columns):
...
dtypes: float32(5)
memory usage: 1.9 GB

I spotted this a little late, but I work with a similar problem (mortgage prepayment models). My solution has been to skip the pandas HDFStore layer and use straight pytables. I save each column as an individual HDF5 array in my final file.
My basic workflow is to first get a CSV file from the database. I gzip it, so it's not as huge. Then I convert that to a row-oriented HDF5 file, by iterating over it in python, converting each row to a real data type, and writing it to a HDF5 file. That takes some tens of minutes, but it doesn't use any memory, since it's only operating row-by-row. Then I "transpose" the row-oriented HDF5 file into a column-oriented HDF5 file.
The table transpose looks like:
def transpose_table(h_in, table_path, h_out, group_name="data", group_path="/"):
# Get a reference to the input data.
tb = h_in.getNode(table_path)
# Create the output group to hold the columns.
grp = h_out.createGroup(group_path, group_name, filters=tables.Filters(complevel=1))
for col_name in tb.colnames:
logger.debug("Processing %s", col_name)
# Get the data.
col_data = tb.col(col_name)
# Create the output array.
arr = h_out.createCArray(grp,
col_name,
tables.Atom.from_dtype(col_data.dtype),
col_data.shape)
# Store the data.
arr[:] = col_data
h_out.flush()
Reading it back in then looks like:
def read_hdf5(hdf5_path, group_path="/data", columns=None):
"""Read a transposed data set from a HDF5 file."""
if isinstance(hdf5_path, tables.file.File):
hf = hdf5_path
else:
hf = tables.openFile(hdf5_path)
grp = hf.getNode(group_path)
if columns is None:
data = [(child.name, child[:]) for child in grp]
else:
data = [(child.name, child[:]) for child in grp if child.name in columns]
# Convert any float32 columns to float64 for processing.
for i in range(len(data)):
name, vec = data[i]
if vec.dtype == np.float32:
data[i] = (name, vec.astype(np.float64))
if not isinstance(hdf5_path, tables.file.File):
hf.close()
return pd.DataFrame.from_items(data)
Now, I generally run this on a machine with a ton of memory, so I may not be careful enough with my memory usage. For example, by default the load operation reads the whole data set.
This generally works for me, but it's a bit clunky, and I can't use the fancy pytables magic.
Edit: The real advantage of this approach, over the array-of-records pytables default, is that I can then load the data into R using h5r, which can't handle tables. Or, at least, I've been unable to get it to load heterogeneous tables.

As noted by others, after some years an 'out-of-core' pandas equivalent has emerged: dask. Though dask is not a drop-in replacement of pandas and all of its functionality it stands out for several reasons:
Dask is a flexible parallel computing library for analytic computing that is optimized for dynamic task scheduling for interactive computational workloads of
“Big Data” collections like parallel arrays, dataframes, and lists that extend common interfaces like NumPy, Pandas, or Python iterators to larger-than-memory or distributed environments and scales from laptops to clusters.
Dask emphasizes the following virtues:
Familiar: Provides parallelized NumPy array and Pandas DataFrame objects
Flexible: Provides a task scheduling interface for more custom workloads and integration with other projects.
Native: Enables distributed computing in Pure Python with access to the PyData stack.
Fast: Operates with low overhead, low latency, and minimal serialization necessary for fast numerical algorithms
Scales up: Runs resiliently on clusters with 1000s of cores Scales down: Trivial to set up and run on a laptop in a single process
Responsive: Designed with interactive computing in mind it provides rapid feedback and diagnostics to aid humans
and to add a simple code sample:
import dask.dataframe as dd
df = dd.read_csv('2015-*-*.csv')
df.groupby(df.user_id).value.mean().compute()
replaces some pandas code like this:
import pandas as pd
df = pd.read_csv('2015-01-01.csv')
df.groupby(df.user_id).value.mean()
and, especially noteworthy, provides through the concurrent.futures interface a general infrastructure for the submission of custom tasks:
from dask.distributed import Client
client = Client('scheduler:port')
futures = []
for fn in filenames:
future = client.submit(load, fn)
futures.append(future)
summary = client.submit(summarize, futures)
summary.result()

It is worth mentioning here Ray as well,
it's a distributed computation framework, that has it's own implementation for pandas in a distributed way.
Just replace the pandas import, and the code should work as is:
# import pandas as pd
import ray.dataframe as pd
# use pd as usual
can read more details here:
https://rise.cs.berkeley.edu/blog/pandas-on-ray/
Update:
the part that handles the pandas distribution, has been extracted to the modin project.
the proper way to use it is now is:
# import pandas as pd
import modin.pandas as pd

One more variation
Many of the operations done in pandas can also be done as a db query (sql, mongo)
Using a RDBMS or mongodb allows you to perform some of the aggregations in the DB Query (which is optimized for large data, and uses cache and indexes efficiently)
Later, you can perform post processing using pandas.
The advantage of this method is that you gain the DB optimizations for working with large data, while still defining the logic in a high level declarative syntax - and not having to deal with the details of deciding what to do in memory and what to do out of core.
And although the query language and pandas are different, it's usually not complicated to translate part of the logic from one to another.

Consider Ruffus if you go the simple path of creating a data pipeline which is broken down into multiple smaller files.

I'd like to point out the Vaex package.
Vaex is a python library for lazy Out-of-Core DataFrames (similar to Pandas), to visualize and explore big tabular datasets. It can calculate statistics such as mean, sum, count, standard deviation etc, on an N-dimensional grid up to a billion (109) objects/rows per second. Visualization is done using histograms, density plots and 3d volume rendering, allowing interactive exploration of big data. Vaex uses memory mapping, zero memory copy policy and lazy computations for best performance (no memory wasted).
Have a look at the documentation: https://vaex.readthedocs.io/en/latest/
The API is very close to the API of pandas.

I recently came across a similar issue. I found simply reading the data in chunks and appending it as I write it in chunks to the same csv works well. My problem was adding a date column based on information in another table, using the value of certain columns as follows. This may help those confused by dask and hdf5 but more familiar with pandas like myself.
def addDateColumn():
"""Adds time to the daily rainfall data. Reads the csv as chunks of 100k
rows at a time and outputs them, appending as needed, to a single csv.
Uses the column of the raster names to get the date.
"""
df = pd.read_csv(pathlist[1]+"CHIRPS_tanz.csv", iterator=True,
chunksize=100000) #read csv file as 100k chunks
'''Do some stuff'''
count = 1 #for indexing item in time list
for chunk in df: #for each 100k rows
newtime = [] #empty list to append repeating times for different rows
toiterate = chunk[chunk.columns[2]] #ID of raster nums to base time
while count <= toiterate.max():
for i in toiterate:
if i ==count:
newtime.append(newyears[count])
count+=1
print "Finished", str(chunknum), "chunks"
chunk["time"] = newtime #create new column in dataframe based on time
outname = "CHIRPS_tanz_time2.csv"
#append each output to same csv, using no header
chunk.to_csv(pathlist[2]+outname, mode='a', header=None, index=None)

The parquet file format is ideal for the use case you described. You can efficiently read in a specific subset of columns with pd.read_parquet(path_to_file, columns=["foo", "bar"])
https://pandas.pydata.org/docs/reference/api/pandas.read_parquet.html

At the moment I am working "like" you, just on a lower scale, which is why I don't have a PoC for my suggestion.
However, I seem to find success in using pickle as caching system and outsourcing execution of various functions into files - executing these files from my commando / main file; For example i use a prepare_use.py to convert object types, split a data set into test, validating and prediction data set.
How does your caching with pickle work?
I use strings in order to access pickle-files that are dynamically created, depending on which parameters and data sets were passed (with that i try to capture and determine if the program was already run, using .shape for data set, dict for passed parameters).
Respecting these measures, i get a String to try to find and read a .pickle-file and can, if found, skip processing time in order to jump to the execution i am working on right now.
Using databases I encountered similar problems, which is why i found joy in using this solution, however - there are many constraints for sure - for example storing huge pickle sets due to redundancy.
Updating a table from before to after a transformation can be done with proper indexing - validating information opens up a whole other book (I tried consolidating crawled rent data and stopped using a database after 2 hours basically - as I would have liked to jump back after every transformation process)
I hope my 2 cents help you in some way.
Greetings.