pandas.DataFrame.convert_dtypes increasing memory usage
Question:
Question to discuss and understand a bit more about pandas.DataFrame.convert_dtypes.
I have this DF imported from a SAS table:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 857613 entries, 0 to 857612
Data columns (total 27 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 cd_unco_tab 857613 non-null object
1 cd_ref_cnv 856389 non-null object
2 cd_cli 849637 non-null object
3 nm_prd 857613 non-null object
4 nm_ctgr_cpr 857613 non-null object
5 ts_cpr 857229 non-null datetime64[ns]
6 ts_cnfc 857613 non-null datetime64[ns]
7 ts_incl 857613 non-null datetime64[ns]
8 vl_cmss_rec 857613 non-null float64
9 qt_prd 857613 non-null float64
10 pc_cmss_rec 857242 non-null float64
11 nm_loja 857242 non-null object
12 vl_brto_cpr 857242 non-null float64
13 vl_cpr 857242 non-null float64
14 qt_dvlc 857613 non-null float64
15 cd_in_evt_espl 857613 non-null float64
16 cd_mm_aa_ref 840959 non-null object
17 nr_est_ctbc_evt 857613 non-null float64
18 nr_est_cnfc_pcr 18963 non-null float64
19 cd_tran_pcr 0 non-null object
20 ts_est 18963 non-null datetime64[ns]
21 tx_est_tran 18963 non-null object
22 vl_tran 18963 non-null float64
23 cd_pcr 0 non-null float64
24 vl_cbac_cli 653563 non-null float64
25 pc_cbac_cli 653563 non-null float64
26 cd_vndr 18963 non-null float64
dtypes: datetime64[ns](4), float64(14), object(9)
memory usage: 176.7+ MB
Basically, the DF is composed of datetime64, float64 and object types. All not memory efficient (as far as I know).
I read a bit about DataFrame.convert_dtypes to optimize memory usage, this is the result:
dfcompras = dfcompras.convert_dtypes(infer_objects=True, convert_string=True, convert_integer=True, convert_boolean=True, convert_floating=True)
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 857613 entries, 0 to 857612
Data columns (total 27 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 cd_unco_tab 857613 non-null string
1 cd_ref_cnv 856389 non-null string
2 cd_cli 849637 non-null string
3 nm_prd 857613 non-null string
4 nm_ctgr_cpr 857613 non-null string
5 ts_cpr 857229 non-null datetime64[ns]
6 ts_cnfc 857613 non-null datetime64[ns]
7 ts_incl 857613 non-null datetime64[ns]
8 vl_cmss_rec 857613 non-null Float64
9 qt_prd 857613 non-null Int64
10 pc_cmss_rec 857242 non-null Float64
11 nm_loja 857242 non-null string
12 vl_brto_cpr 857242 non-null Float64
13 vl_cpr 857242 non-null Float64
14 qt_dvlc 857613 non-null Int64
15 cd_in_evt_espl 857613 non-null Int64
16 cd_mm_aa_ref 840959 non-null string
17 nr_est_ctbc_evt 857613 non-null Int64
18 nr_est_cnfc_pcr 18963 non-null Int64
19 cd_tran_pcr 0 non-null Int64
20 ts_est 18963 non-null datetime64[ns]
21 tx_est_tran 18963 non-null string
22 vl_tran 18963 non-null Float64
23 cd_pcr 0 non-null Int64
24 vl_cbac_cli 653563 non-null Float64
25 pc_cbac_cli 653563 non-null Float64
26 cd_vndr 18963 non-null Int64
dtypes: Float64(7), Int64(8), datetime64[ns](4), string(8)
memory usage: 188.9 MB
Most columns were changed from object to strings and float64 to int64, so, it would reduce memory usage, but as we can see, the memory usage increased!
Any guess?
Answers:
After doing some analysis it seems like there is an additional memory overhead while using the new Int64/Float64 Nullable dtypes. Int64/Float64 dtypes takes approximately 9 bytes while int64/float64 dtypes takes 8 bytes to store a single value.
Here is a small example to demonstrate this:
pd.DataFrame({'col': range(10)}).astype('float64').memory_usage()
Index 128
col 80 # 8 byte per item * 10 items
dtype: int64
pd.DataFrame({'col': range(10)}).astype('Float64').memory_usage()
Index 128
col 90 # 9 byte per item * 10 items
dtype: int64
Now, coming back to your example. After executing convert_dtypes around 15 columns got converted from float64 to Int64/Float64 dtypes. Now lets calculate the amount of extra bytes required to store the data with new types. The formula would be fairly simple: n_columns * n_rows * overhead_in_bytes
>>> extra_bytes = 15 * 857613 * 1
>>> extra_mega_bytes = extra_bytes / 1024 ** 2
>>> extra_mega_bytes
12.2682523727417
Turns out extra_mega_bytes is around 12.26 MB which is approximately same as the difference between the memory usage of your new and old dataframe.
Some details about new nullable integer datatype:
Int64/Float64(notice the first capital letter) are some of the new nullable types that are introduced for the first time in pandas version>=0.24 on a high level they allow you use pd.NA instead of pd.NaN/np.nan to represent missing values and implication of this can be better understood in the following example:
s = pd.Series([1, 2, np.nan])
print(s)
0 1.0
1 2.0
2 NaN
dtype: float64
Let’s say you have a series s now when you check the dtype, pandas will automatically cast it to float64 because of presence of null values this is not problematic in most of cases but in case you have an column which acts as an identifier the automatic conversion to float is undesirable. To prevent this pandas has introduced these new nullable integer type.
s = pd.Series([1, 2, np.nan], dtype='Int64')
print(s)
0 1
1 2
2 <NA>
dtype: Int64
Some details on string dtype
As of now there isn’t a much performance and memory difference when using the new string type but this can change in the near future. See the quote from pandas docs:
Currently, the performance of object dtype arrays of strings and
StringArray are about the same. We expect future enhancements to
significantly increase the performance and lower the memory overhead
of StringArray.
Question to discuss and understand a bit more about pandas.DataFrame.convert_dtypes.
I have this DF imported from a SAS table:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 857613 entries, 0 to 857612
Data columns (total 27 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 cd_unco_tab 857613 non-null object
1 cd_ref_cnv 856389 non-null object
2 cd_cli 849637 non-null object
3 nm_prd 857613 non-null object
4 nm_ctgr_cpr 857613 non-null object
5 ts_cpr 857229 non-null datetime64[ns]
6 ts_cnfc 857613 non-null datetime64[ns]
7 ts_incl 857613 non-null datetime64[ns]
8 vl_cmss_rec 857613 non-null float64
9 qt_prd 857613 non-null float64
10 pc_cmss_rec 857242 non-null float64
11 nm_loja 857242 non-null object
12 vl_brto_cpr 857242 non-null float64
13 vl_cpr 857242 non-null float64
14 qt_dvlc 857613 non-null float64
15 cd_in_evt_espl 857613 non-null float64
16 cd_mm_aa_ref 840959 non-null object
17 nr_est_ctbc_evt 857613 non-null float64
18 nr_est_cnfc_pcr 18963 non-null float64
19 cd_tran_pcr 0 non-null object
20 ts_est 18963 non-null datetime64[ns]
21 tx_est_tran 18963 non-null object
22 vl_tran 18963 non-null float64
23 cd_pcr 0 non-null float64
24 vl_cbac_cli 653563 non-null float64
25 pc_cbac_cli 653563 non-null float64
26 cd_vndr 18963 non-null float64
dtypes: datetime64[ns](4), float64(14), object(9)
memory usage: 176.7+ MB
Basically, the DF is composed of datetime64, float64 and object types. All not memory efficient (as far as I know).
I read a bit about DataFrame.convert_dtypes to optimize memory usage, this is the result:
dfcompras = dfcompras.convert_dtypes(infer_objects=True, convert_string=True, convert_integer=True, convert_boolean=True, convert_floating=True)
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 857613 entries, 0 to 857612
Data columns (total 27 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 cd_unco_tab 857613 non-null string
1 cd_ref_cnv 856389 non-null string
2 cd_cli 849637 non-null string
3 nm_prd 857613 non-null string
4 nm_ctgr_cpr 857613 non-null string
5 ts_cpr 857229 non-null datetime64[ns]
6 ts_cnfc 857613 non-null datetime64[ns]
7 ts_incl 857613 non-null datetime64[ns]
8 vl_cmss_rec 857613 non-null Float64
9 qt_prd 857613 non-null Int64
10 pc_cmss_rec 857242 non-null Float64
11 nm_loja 857242 non-null string
12 vl_brto_cpr 857242 non-null Float64
13 vl_cpr 857242 non-null Float64
14 qt_dvlc 857613 non-null Int64
15 cd_in_evt_espl 857613 non-null Int64
16 cd_mm_aa_ref 840959 non-null string
17 nr_est_ctbc_evt 857613 non-null Int64
18 nr_est_cnfc_pcr 18963 non-null Int64
19 cd_tran_pcr 0 non-null Int64
20 ts_est 18963 non-null datetime64[ns]
21 tx_est_tran 18963 non-null string
22 vl_tran 18963 non-null Float64
23 cd_pcr 0 non-null Int64
24 vl_cbac_cli 653563 non-null Float64
25 pc_cbac_cli 653563 non-null Float64
26 cd_vndr 18963 non-null Int64
dtypes: Float64(7), Int64(8), datetime64[ns](4), string(8)
memory usage: 188.9 MB
Most columns were changed from object to strings and float64 to int64, so, it would reduce memory usage, but as we can see, the memory usage increased!
Any guess?
After doing some analysis it seems like there is an additional memory overhead while using the new Int64/Float64 Nullable dtypes. Int64/Float64 dtypes takes approximately 9 bytes while int64/float64 dtypes takes 8 bytes to store a single value.
Here is a small example to demonstrate this:
pd.DataFrame({'col': range(10)}).astype('float64').memory_usage()
Index 128
col 80 # 8 byte per item * 10 items
dtype: int64
pd.DataFrame({'col': range(10)}).astype('Float64').memory_usage()
Index 128
col 90 # 9 byte per item * 10 items
dtype: int64
Now, coming back to your example. After executing convert_dtypes around 15 columns got converted from float64 to Int64/Float64 dtypes. Now lets calculate the amount of extra bytes required to store the data with new types. The formula would be fairly simple: n_columns * n_rows * overhead_in_bytes
>>> extra_bytes = 15 * 857613 * 1
>>> extra_mega_bytes = extra_bytes / 1024 ** 2
>>> extra_mega_bytes
12.2682523727417
Turns out extra_mega_bytes is around 12.26 MB which is approximately same as the difference between the memory usage of your new and old dataframe.
Some details about new nullable integer datatype:
Int64/Float64(notice the first capital letter) are some of the new nullable types that are introduced for the first time in pandas version>=0.24 on a high level they allow you use pd.NA instead of pd.NaN/np.nan to represent missing values and implication of this can be better understood in the following example:
s = pd.Series([1, 2, np.nan])
print(s)
0 1.0
1 2.0
2 NaN
dtype: float64
Let’s say you have a series s now when you check the dtype, pandas will automatically cast it to float64 because of presence of null values this is not problematic in most of cases but in case you have an column which acts as an identifier the automatic conversion to float is undesirable. To prevent this pandas has introduced these new nullable integer type.
s = pd.Series([1, 2, np.nan], dtype='Int64')
print(s)
0 1
1 2
2 <NA>
dtype: Int64
Some details on string dtype
As of now there isn’t a much performance and memory difference when using the new string type but this can change in the near future. See the quote from pandas docs:
Currently, the performance of object dtype arrays of strings and
StringArrayare about the same. We expect future enhancements to
significantly increase the performance and lower the memory overhead
of StringArray.