After using cache() in pyspark the row count is wrong - dataframe

Question

When I perform a row count operation on my dataframe with/without cache(), I get a different result.
The example is the following:
videosDF = videosRawDF.withColumn("trending_date", F.to_date("trending_date", "yy.dd.MM"))\
.withColumn("publish_time", F.from_unixtime(F.unix_timestamp('publish_time', "yyyy-MM-dd'T'HH:mm:ss")).cast("timestamp"))\
.dropna(subset = ["trending_date", "publish_time"])
videosDF.count()
120746
On the other hand if I only add .cache() the result changes
videosDF = videosRawDF.withColumn("trending_date", F.to_date("trending_date", "yy.dd.MM"))\
.withColumn("publish_time", F.from_unixtime(F.unix_timestamp('publish_time', "yyyy-MM-dd'T'HH:mm:ss")).cast("timestamp"))\
.dropna(subset = ["trending_date", "publish_time"])
videosDF.cache()
videosDF.count()
0
The correct solution after performing the drop is 120746 rows, but I am forced to perform the cache. What I can do?
Thank you for reading.
EDIT:
The dataset can be found in: https://www.kaggle.com/datasets/datasnaek/youtube-new
and my code until now is the following:
from pyspark.sql.types import StructType, StructField, StringType, IntegerType,BooleanType
from pyspark.sql import functions as F
ruta = "gs://ucmbucketrafael/data/"
youtubeSchema = StructType([
StructField('video_id', StringType(), nullable = True),
StructField('trending_date', StringType(), nullable = True),
StructField('title', StringType(), nullable = True),
StructField('channel_title', StringType(), nullable = True),
StructField('category_id', StringType(), nullable = True),
StructField('publish_time', StringType(), nullable = True),
StructField('tags', StringType(), nullable = True),
StructField('views', IntegerType(), nullable = True),
StructField('likes', IntegerType(), nullable = True),
StructField('dislikes', IntegerType(), nullable = True),
StructField('comment_count', IntegerType(), nullable = True),
StructField('comments_disabled', BooleanType(), nullable = True),
StructField('ratings_disabled', BooleanType(), nullable = True),
StructField('video_error_or_removed', BooleanType(), nullable = True),
StructField('description', StringType(), nullable = True)
])
USvideosDF = spark.read.schema(youtubeSchema)\
.option("header", "true")\
.option("quote", "\"").option("escape", "\"")\
.option("mode","DROPMALFORMED")\
.csv(ruta+"youtube_USvideos.csv")\
.withColumn("pais",F.lit("EEUU"))\
.drop("description")
CAvideosDF = spark.read.schema(youtubeSchema)\
.option("header", "true")\
.option("quote", "\"").option("escape", "\"")\
.option("mode","DROPMALFORMED")\
.csv(ruta+"youtube_CAvideos.csv")\
.withColumn("pais",F.lit("CA"))\
.drop("description")
GBvideosDF = spark.read.schema(youtubeSchema)\
.option("header", "true")\
.option("quote", "\"").option("escape", "\"")\
.option("mode","DROPMALFORMED")\
.csv(ruta+"youtube_GBvideos.csv")\
.withColumn("pais",F.lit("GB"))\
.drop("description")
videosRawDF = USvideosDF.union(CAvideosDF).union(GBvideosDF)
videosDF = videosRawDF.withColumn("trending_date", F.to_date("trending_date", "yy.dd.MM"))\
.withColumn("publish_time", F.from_unixtime(F.unix_timestamp('publish_time', "yyyy-MM-dd'T'HH:mm:ss")).cast("timestamp"))\
.dropna(subset = ["trending_date", "publish_time"])
videosDF.cache()
videosDF.count()

Answer 1

I have already discovered the problem. It is the interaction between DROPMALFORMED and the scheme. If the schema is deleted when reading the csv it already works correctly.